1 Introduction
An important step in the visual processing hierarchy is putting together fragments of features into coherent objects and inferring the spatial relationship between them. The feature fragments can be based on color, orientation, texture, etc. Grouping Wagemans et al. (2012a, b) refers to the mechanism by which the feature fragments are put together to form perceptual objects. Such objects in the real world may be isolated, fully occluding one another or partially occluding, depending on the observer’s viewpoint. In the context of partially occluding objects, Figureground organization (FGO) refers to determining which side of an occlusion boundary is the occluder, closer to the observer, referred to as figure and which side is the occluded, far away from the observer, termed as ground.
Gestalt psychologists have identified a variety of cues that mediate the process of FGO Koffka (1935)
. Based on the spatial extent of information integration, these cues can be classified into local and global cues. Global cues such as symmetry
Bahnsen (1928), surroundedness Palmer (1999), and size Fowlkes et al. (2007) of regions integrate information over a large spatial extent to determine figure/ground relationship between objects. Local cues, on the other hand, achieve the same by analysis of only a small neighborhood near the boundary of an object. Hence, they are particularly attractive from a computational standpoint. Some examples of local cues are Tjunctions Heitger et al. (1992) and shading Huggins et al. (2001), including extremal edges Palmer and Ghose (2008); Ramenahalli et al. (2011).The neural mechanism by which FGO is achieved in the visual cortex is an active area of research, referred to as Border Ownership (BO) coding. The contour fragments forming an object’s boundary are detected by Simple and Complex cells in the area V1 of primate visual cortex with their highly localized, retinotopically organized receptive fields. Cells in area V2, which receive input from V1 Complex cells, were found to code for BO by preferentially firing at a higher rate when the figural object was located on the preferred side of the BO coding neuron at its preferred orientation, irrespective of local contrast
Zhou et al. (2000). Recently, Williford and von der Heydt (2016), remarkably show for the first time, that V2 neurons maintain the same BO preference properties even for objects in complex natural scenes.Many computational models Craft et al. (2007); Roelfsema et al. (2002); Zhaoping (2005) have been proposed to explain the neural mechanism by which FGO or BO coding is achieved in the visual cortex. Based on the connection mechanism, those models can be classified as feedforward, feedback Mihalas et al. (2011); Hu et al. (2019) or lateral interaction models Zhaoping (2005). In this work, we present a neurally motivated, feedforward computational model of FGO incorporating both local and global cues. While we do not attempt to exactly mimic the neural processing at every step, we attempt to keep it as biologically motivated as possible.
The FGO model we develop has three independent feature channels, Color, Intensity and Orientation. The main computational construct of the model is a BO computation mechanism that embodies Gestalt principles of convexity, surroundedness and parallelism, which is identical to all feature channels. In addition, we introduce many additional modifications to make it suitable for performing FGO and to incorporate local cues, as detailed in Section 3. The model, applicable to any natural image, is tested on the widely used BSDS figure/ground dataset. First, we show that even the model with only global cues, devoid of any local cues achieves good results on the BSDS figure/ground dataset. Let us call this the Reference model, against which we compare the performance of models with added local cues.
We add two local cues to the reference model, Spectral Anisotropy Ramenahalli et al. (2014) and TJunctions. The motivation behind adding local cues is their relatively low computational cost compared to global cues. Spectral Anisotropy (SA) was shown to be a valid cue for FGO Ramenahalli et al. (2014, 2011, 2012) in predicting which side of an occlusion boundary is figure and which the background. Moreover, SA can be computed efficiently in a biologically plausible (See Section 4.1) manner using convolutions, making it an attractive candidate. TJunctions are commonly viewed as one of the strongest cues of occlusion and their computation can be explained on the basis of endstopped cells Heitger et al. (1992); Heitger and von der Heydt (1993); Hansen and Neumann (2002). This is the biological motivation to incorporate them into the model.
We have only a few FGO cues, specifically two local cues in our model. Both local cues influence the Orientation channel only as the properties they capture are more closely related to this feature. Certainly, many more local cues and global cues would be needed for best performance in real world images. But, here our primary motivation is to develop a common computational framework and investigate how these local and global cues can be incorporated into a model of FGO. Second, our purpose is to verify whether local cues can coexist along with the global cues. If so, how useful are these local cues? Can they lead to a statistically significant improvement in the model’s performance when added alone? Finally, are these local cues mutually facilitatory leading to even further improvement, when added together? For these purposes, the minimalistic model with few global cues and even fewer local cues added to only one of the three feature channels provides an excellent analysis framework. Our goal is to study, from first principles, the effect of local and global cues in FGO, not necessarily to build a model with best performance. However, we compare the performance of our model with state of the art models of FGO, which are not biologically motivated, and show that our model performs competitively.
2 Related Work
FGO has been an active area of research in Psychology since nearly a century Rubin (1921) ago. The Gestalt principles of FGO and grouping such as common fate, symmetry, good continuation, similarity etc were formulated by Max Wertheimer Wertheimer (1923) along with Kurt Koffka Koffka (1935) and many others. Excellent reviews about the Gestalt principles of FGO and grouping can be found in Wagemans et al. (2012a, b). It is an active area of research in neuroscience Lamme (1995); Super and Lamme (2007); Zhou et al. (2000); Williford and von der Heydt (2014)
and computer vision
Ren et al. (2006); Hoiem et al. (2011); Teo et al. (2015) as well. We limit our literature review to computational models only. Even though the terms “FGO”, “BO” or “grouping” are not used in many publications we reviewed, the common goal in all of them is related to inferring depth ordering of objects.A local shapeme based model employing Conditional Random Fields (CRF) to enforce global consistency at Tjunctions was proposed in Ren et al. (2006). Hoiem et al. Hoiem et al. (2007, 2011) used a variety of local region, boundary, Gestalt and depth based cues in a CRF model to enforce consistency between boundary and surface labels. An optimization framework Amer et al. (2010) to obtain a 2.1D sketch by constraining the “hat” of the Tjunction to be figure and “stem” to be ground was proposed, which uses human labeled contours and Tjunctions. In an extension Amer et al. (2015), a reformulated optimization over regions, instead of pixels, was proposed. By using various cues such as, curve and junction potentials, convexity, lowerregion, fold/cut and parallelism, Leichter and Lindenbaum (2009) train a CRF model to enforce global consistency. In a series of papers Palou and Salembier (2013, 2012, 2011) show how image segmentation and depth ordering (FGO) can be performed using only lowlevel cues. Their model uses Binary Partition Trees (BPT) Salembier and Garrido (2000) for hierarchically representing regions of an image, performs depth ordering by iteratively pruning the branches of BPT enforcing constraints based on Tjunctions and other depth related cues. In a recent work Teo et al. (2015)
, which uses Structured Random Forests (SRF) for boundary detection, simultaneous boundary detection and figureground labeling is performed. They use shape cues, extremal edge basis functions
Ramenahalli et al. (2011), closure, image torque Nishigaki et al. (2012) etc to train the SRFs.Yu et al. (2001) present a hierarchical Markov Random Field (MRF) model incorporating rules for continuity of depth on surfaces, discontinuity at edges between surfaces and local cues such as T and Ljunctions. The model learns from a couple examples and effectively does depth segregation, thereby FGO. In Baek and Sajda (2005)
, a neurally plausible model integrating multiple figureground cues using belief propagation in Bayesian networks with leaky integrate and fire neurons was proposed. A simultaneous segmentation and figureground labeling algorithm was reported in
Maire (2010) which uses Angular Embedding Yu (2009) to influence segmentation cues from figureground cues and viceversa. Similar attempts with primary goal of segmenting images and labeling object classes using figureground cues can be seen in Ion et al. (2011, 2014).Differentiation/Integration for Surface Completion (DISC) model Kogo et al. (2010) was proposed in which BO is computed by detecting local occlusion cues such as T and L junctions and comparing nonjunction border locations with junction locations for BO consistency with the local cues. A Bayesian belief network based model was proposed Froyen et al. (2010) in which local cues (curvature and Tjunctions) interact with medial axis or skeleton of the shape to determine BO.
In one of the early attempts Kienker et al. (1986), a two layer network with connections between “computational units” within and across layers is proposed. These units integrate bottomup edge input with topdown attention input to realize FGO. Grossberg and Mingolla (1985); Grossberg (1994) propose that a reciprocal interaction between a Boundary Contour System (BCS) extracting edges and a Feature Contour System (FCS) extracting surfaces achieves not only FGO, but also 3D perception. A model of contour grouping and FGO was proposed in Heitger and von der Heydt (1993) central to which is a “grouping” mechanism. The model not only generates figureground labels, but also simulates the perception of illusory contours. Another influential model was proposed in Sajda and Finkel (1995) with feedback and feedforward connections having 8 different computational modules to obtain representations of contours, surfaces and depth. Roelfsema et al. (2002); Jehee et al. (2007) propose multilayer feedback networks resembling the neural connection pattern in the visual cortex to perform BO assignment through feedback from higher areas. Li Zhaoping et al. Li (2003, 2000) propose a model of FGO based on V1 mechanisms. The model consists of orientation selective V1 neurons which influence surrounding neurons through monosynaptic excitatory and disynaptic suppressive connections. The excitatory lateral connections mimic colinear excitation Kapadia et al. (1995) and crossorientation facilitation Slllito et al. (1995), while inhibitory connections model the isoorientation suppression Knierim and Van Essen (1992). In a related model Zhaoping (2005), neurons in V2 having properties of convexity preference, good continuation and proximity was presented. A BO coding model which detects curvatures, LJunctions and sends proportional signals to a BO layer was proposed by Kikuchi and Akashi (2001), where BO signals are propagated along the contour for two sides of BO.
The model proposed by Craft et al. (2007) consists of edge selective cells, BO cells and multiscale grouping (G) cells. The G cells send excitatory feedback to those BO cells that are cocircular and point to the center of the annular G cell receptive field. The model incorporates Gestalt principles of convexity, proximity and closure. But, it is a feedback model tested only on simple geometric shapes, not realworld natural images. Several models Mihalas et al. (2011); Russell et al. (2014); Molin et al. (2013); Hu and Niebur (2017) with similar computational mechanisms have been proposed to explain various phenomena related to FGO, saliency, spatial attention, etc. A model akin to Craft et al. (2007) was proposed in Layton et al. (2012), where in addition to G cells the model consists of region cells at multiple scales. In a feedback model Domijan and Šetić (2008) based on the interaction between dorsal and ventral streams, surfaces which are of smaller size, greater contrast, convex, closed, having higher spatial frequency are preferentially determined as figures. The model also accounts for figureground cues such as lower region and topbottom polarity. In a series of papers Sakai et al. (2012); Nishimura and Sakai (2004, 2005) Sakai and colleagues formulate a BO model in which localized, asymmetric surround modulation is used to detect contrast imbalance, which then leads to FGO. Russell et al. (2014) propose a feedforward model with Grouping and Border Ownership cells to study protoobject based saliency. Though our model is inspired by this work, the goal of Russell et al. (2014) model is to explain the formation of protoobjects Rensink (2000) and saliency prediction, not FigureGround Organization. Another related model is proposed by Hu et al. (2019), which is a recurrent model with feedback connections, devoid of any local cues. To the best of our knowledge our’s is the first feedforward model of FGO with both local and global cues. Also, this the the first such model tested on realworld images of the BSDS300 figureground dataset commonly used as a benchmark for FGO in natural images.
3 Model Description
The model consists of three independent features channels, Color, Intensity and Orientation. The features are computed at multiple scales to achieve scale invariance. Orientation selective V1 Simple and Complex cells Adelson and Bergen (1985) are excited by edge fragments of objects within their receptive field (Figure 1). Let us denote the contrast invariant response of a Complex cell at location by , where is the preferred orientation of the cell and is the spatial frequency. As the spatial frequency (see Table 2 for all parameters of the model) is same of all edge responsive cells in our model, except when explicitly stated otherwise (Section 4.1), we omit this variable for the most part. Each active Complex cell, activates a pair of BO cells, one with a BO preference direction, (a counterclockwise rotation with respect to ) denoted as , and the other with BO preference, denoted as . When we talk about the BO response related to a specific figure/ground cue, be it local or global, a subscript is added to the right of the variable. For example, would be used to denote the BO response related to TJunctions. Likewise, when specifying scale is necessary, it is denoted by superscript, . For example, denotes Complex cell response for orientation at location, and scale, . On the other hand, when we need to explicitly specify the feature we talk about, a subscript is added to the left of the variable. For example, represents the BO response for the Color feature channel. When a specific BO direction, feature, cue, scale or a location is not important, we just refer to them as, cells, cells, etc. Same applies in all such situations.
Without the influence of any local or global cues, the responses of both BO cells at a location will be equal, hence the figure direction at that location is arbitrary. The centersurround cells, denoted as cells, bring about global scene context integration by modulating the cell activity. The cells (Figure 1) extract light objects on dark background, while cells code for dark objects on light background. Without the influence of local cues, this architecture embodies the Gestalt properties of convexity, surroundedness and parallelism (global cues).
The local cues (see Section 4 for computational details of local cues) modulate cell activity additionally. Similar to cells, a pair of Spectral Anisotropy cells exist for the two opposite BO preference directions at each location, which capture local texture and shading gradients (see Section 4.1 for SA computation) on the two sides of the border. Let us denote by the cell capturing Spectral Anisotropy for BO direction, likewise for the opposite BO direction. The TJunction cells (see Section 4.2 for computational details) also come in pairs, for the two opposite BO directions. Similar to cells, hold the TJunction cue information for the two antagonistic BO directions, . Both these type of cells excite cells of the same BO direction and inhibit the opposite BO direction cells. For example, excites and inhibits .
The influence of cells, cells and cells on cells is controlled by a set of weights (not shown in Figure 1). Local cues are active in the Orientation channel only. The interplay of all these cues leads to the emergence of figure/ground relations strongly biased for one of the two BO directions at each location. The network architecture depicted in Figure 1 is the same computational construct that is applied at every scale, for every feature and orientation. The successive stages of model computation are explained in the following subsections.
3.1 Computation of feature channels
We consider Color, Intensity and Orientation as three independent feature channels in our model, the computation of each is described in the following sections.
3.1.1 Intensity channel
The input image consists of Red (), Blue () and Green () color channels. The intensity channel, is computed as average of the three channels, . As with all other feature channels, a multiresolution image pyramid is constructed from the intensity channel (Section 3.2). The multiresolution analysis allows incorporation of scale invariance into the model.
3.1.2 Color opponency channels
The color channels are first normalized by dividing each , or value by . From the normalized , , channels, four color channels, Red (), Green (), Blue () and Yellow () are computed as,
(1) 
(2) 
(3) 
(4) 
In Eq 4, the symbol, denotes absolute value.
The four opponent color channels, , , and are computed as,
(5) 
(6) 
(7) 
(8) 
3.1.3 Orientation channel
The Orientation channel is computed using the canonical model of visual cortex Adelson and Bergen (1985)
, where quadrature phase, orientation selective, Gabor kernels are used to model the V1 simple cells. The responses of Simple cells are nonlinearly combined to obtain the contrast invariant, orientation selective response of the Complex cell. Mathematically, the receptive fields of even and odd symmetric Simple cells can be modeled as the cosine and sine components of a complex Gabor function  a sinusoidal carrier multiplied by a Gaussian envelope. The RF of a Simple Even cell,
is given by,(9) 
where, and are the rotated coordinates,
is the standard deviation of the Gaussian envelope,
is the spatial aspect ratio (controlling how elongated or circular the filter profile is), is the spatial frequency of the cell and is the preferred orientation of the simple cell. Similarly, the receptive field of a Simple Odd cell is defined as,(10) 
Simple even and odd cells responses, respectively denoted and are computed by correlating the intensity image, with the respective RF profiles. The Complex cell response, is calculated as,
(11) 
Eight orientations in the range, , at intervals of are used.
3.2 Multiscale pyramid decomposition
Let us denote a feature map, be it Orientation (), Color (, , etc) or Intensity feature map, at image resolution by a common variable, . The next scale feature map, is computed by downsampling . The downsampling factor can be either
(halfoctave) or 2 (full octave). Bilinear interpolation is used to compute values in the downsampled feature map,
, which is the same interpolation scheme used in all cases of up/down sampling. Similarly, any feature map of a lower scale, is computed by downsampling the higher scale feature map, by the appropriate downsampling factor. As the numerical value of increases, the resolution of the map at that level in the pyramid decreases. The feature pyramids thus obtained are used to compute BO pyramids explained the next section.In addition to the multiscale pyramids of independent feature channels, we compute the multiscale local cue pyramids for SA and TJunctions as well. To denote the local cue map at a specific scale, as with feature pyramids, the scale parameter is used. For example, denotes the Spectral Anisotropy feature map for border ownership direction at scale, . Similarly TJunction pyramids at different scales for BO directions are denoted by . The local cue pyramids are computed by successively downsampling the local cue maps at native resolution, and (see Section 4 for their computation details).
3.3 Border Ownership pyramid computation
The operations performed on any of the features ( or ) or the subtype of features like is the same. BO responses are computed by modulating by the activity of centersurround feature differences on either sides of the border. Each feature map, , is correlated with the centersurround filters to get centersurround () difference feature pyramids. Two types of centersurround filters, (ONcenter) and are defined as,
(12) 
(13) 
where are the standard deviations of the outer and inner Gaussian kernels respectively.
The centersurround dark pyramid, is obtained by correlating the feature maps, with the filter followed by halfwave rectification,
(14) 
which detects weak/dark features surrounded by strong/light ones. In Eq 14, the symbol, denotes 2D correlation MATLAB (9 30). Similarly, to detect strong features surrounded by weak background, a pyramid is computed as,
(15) 
The pyramid computation is performed this way for all feature channels except for the Orientation channel. For the Orientation feature channel, feature contrasts are not typically symmetric as in the case of other features, but oriented at a specific angle. Hence, the and filter kernels in Equations 14 and 15 are replaced by even symmetric Gabor filters, (ONcenter) and (OFFcenter) of opposite polarity respectively. But, in this case, different set of parameter values are used. Instead of , and used in Section 3.1.3, here we use , and respectively. The parameter values are modified in this case such that the width of the center lobe of the even Gabor filters (ON and OFFcenter) matches the zero crossing diameter of the and filter kernels in Equations 14 and 15. As a result, the ONcenter Gabor kernel detects bright oriented edges in a dark background, instead of symmetric feature discontinuities detected by . Similarly, the OFFcenter Gabor filter detects activity of dark edges on bright backgrounds.
An important step in BO computation is normalization of the centersurround feature pyramids, and . Let be used to denote the normalization operation, which is same as the normalization used in Itti et al. (1998), but done after rescaling and pyramids to have the same range, . Similarly the local cue pyramids, and are also normalized using the same method and in the same range, . In the same way, pyramids are also normalized. This normalization step enables comparison of different features and local cues on the same scale, hence the combination of feature and local cue pyramids.
Since, we compute BO on the normalized light and dark CS pyramids, and separately and combine them at a later stage, let us denote, the corresponding BO pyramids by and respectively. We explain the BO pyramid computation for and which have a BO preference direction of . Computation of and is analogous.
Let denote the kernel responsible for mapping the object activity from normalized and
pyramids to the object edges, which is implemented with von Mises distribution. von Mises distribution is a normal distribution on a circle
Weisstein (9 30). The unnormalized von Mises distribution, is defined as Russell et al. (2014),(16) 
where pixels is the radius of the circle on which the von Mises distribution is defined, is the angle at which the normal distribution is concentrated Weisstein (9 30) on the circle (also called mean direction), and is the modified Bessel function of the first kind. The distribution is then normalized as,
(17) 
is computed analogously.
The BO pyramid, for light objects on dark background capturing the BO activity for direction is computed as,
(18) 
Similarly, the BO pyramid for direction for a dark object on light background is obtained by correlating normalized maps with and summing the responses for all scales greater than the scale, at which BO map is being computed as,
(19) 
where, is the synaptic weight for the inhibitory signal from the feature map of opposite contrast polarity. The symbol, is used to denote pixelwise addition of responses from all scales greater than , by first upsampling the response to the scale at which is being computed. The other two pyramids, and for the opposite BO direction are computed analogously.
With the BO pyramids related to dark and light pyramids already computed, we turn our attention to the computation of the local cue related BO pyramids. The local cue pyramids at different scales, and are constructed, as explained in Sections 4.1 and 4.2, by successively downsampling the local cue maps computed at native image resolution. Both local cues excite cells of the same BO direction and inhibit the opposite BO direction cells.
The BO pyramid for BO direction related to the local cue, SA denoted as, is computed as,
(20) 
where we can see the SA cell () having same BO preference as the BO cell, has an excitatory effect on the BO cell, but has an inhibitory effect. The synaptic weight, remains unchanged as in Eqs 18 and 19. The BO pyramid, related to SA, for opposite BO direction is computed in the same way.
The BO pyramid related to TJunctions for the BO direction, is computed as,
(21) 
The corresponding TJunction pyramid for the opposite BO direction, , denoted as is computed analogously.
The combined BO pyramid for direction, is computed by summing global and local cue specific BO pyramids as,
(22) 
where , and are weights such that , that control the contribution of , and cues to the BO response at that location respectively. By setting the weights to 0 or 1, we can study the effect of individual cue on BO response. It should be noted that the local cues are active only for the Orientation channel, so for the other channels, and will be set to zero, by default. In the absence of local cues, combination of light and dark BO pyramids (first term in Eq 22) results in contrast polarity invariant BO response. The corresponding BO pyramid for opposite BO preference, is computed as in Eq 22 by summing the light, dark and local cue BO pyramids of opposite BO preference.
Since the BO responses, , are computed for each orientation, there will be multiple BO responses active at a given pixel location. But the boundary between figure and ground can only belong to the figure side, i.e. there can only be one winning BO response for a given location. So, the winning BO response, denoted as is computed as,
(23) 
where is the orientation for which absolute difference between antagonistic pair of BO responses is maximum over all orientations. This gives the edge orientation at that location. So, the winning BO pyramid, has nonzero response at a location only if the difference between the corresponding pair of BO responses for is nonnegative. The winning BO pyramid, for the opposite direction is computed analogously.
Upto this point, the computation for all feature channels is identical. Now, if we denote the feature specific winning BO pyramid for direction for the Color channel by , Intensity feature channel by and Orientation feature channel by , then the final BO map, for BO direction is computed by linearly combining the upsampled feature specific BO maps across scales as,
(24) 
where represents pixelwise addition of feature specific BO responses across scales after upsampling each map to native resolution of the image. Similarly, is computed for BO direction. As we can see in Eq 24, the contribution of every feature channel to the final BO map is the same, i.e. , feature combination is equally weighted. Ten spatial scales () are used. All parameters of the model are summarized in Table 2. In the end, we get 16 BO maps at image resolution, 8 each for and BO directions respectively.
4 Adding local cues
Both local cues, SA and TJunctions are computed at the native resolution of the images, but they influence BO cells of all scales as described in Eqs. 20, 21. In other words, the cues are computed once based on the analysis local image neighborhood, but their effect is not local^{1}^{1}1Should the effect of local cues also be local? See Section 7 for related discussion.
4.1 Computation of Spectral Anisotropy
Spectral Anisotropy, a local cue for FGO, that captures intensity and texture gradients very close to object boundaries, is computed by pooling Complex cell responses of various spatial frequencies from small image regions on either sides of the boundary (Figure 2). This computation is neurally/biologically plausible.
SA at any location, in the image, for a specific orientation, and for the BO direction, is computed for one side of the border (side determined by the BO direction, ) as,
(25) 
where . The Complex cell response, , is computed as explained in Section 3.1.3, but with a different set of parameters, , , instead of , and respectively. The values of , , and other relevant parameters are listed in Table 1. Filter size is equal to and is the perpendicular distance between the point, at which SA is being computed and the center of the Gabor filters. The centers of even and odd symmetric Gabor filters, hence the Complex cell are all located at , from where the complex cell responses are pooled to compute SA. The term, determines the number of lobes in the Gabor filters. It is 2 or 4 for even symmetric Simple cells and 3 or 5 for odd symmetric Simple cells. The location from which Complex cell responses are pooled, is computed as,
(26) 
(27) 
Similarly, SA at the same location, , but for the opposite side of border at the same orientation, is computed as,
(28) 
where,
(29) 
(30) 
So, for every location there will be two SA cells capturing the spatial intensity and texture gradients on the two sides abutting the border. It has to be noted that the major axis orientation of the Gabor filters is the same as the local border orientation, . This is because we want to capture the variation of spectral power in a direction orthogonal to the object boundary, which is captured by the Complex cells with their orientation parallel to the object boundary. This biologically plausible computation of SA with Complex cells responses captures the anisotropic distribution of high frequency spectral power on figure side we observed in Ramenahalli et al. (2014). The SA maps thus obtained are decomposed into multiscale pyramids, , where superscript, denotes scale, by successive downsampling, which are used to compute the cue specific BO pyramids as explained in Section 3.3, Equation 20.
4.2 Detecting of TJunctions
The object edges and the regions bound by those edges called “segments” are obtained using the gPb+ucm+OWT image segmentation algorithm Arbelaez et al. (2011), referred to as the gPb algorithm in other parts of this work. Image segmentation, partitioning of an image into disjoint regions, is considered as a preprocessing step occurring prior to FGO. The edges obtained using the gPb algorithm are represented as a Contour Map as shown in Figure 3 (B). The corresponding Segmentation Map is shown in Figure 3 (C). The Contour Map has uniquely numbered pieces of contours that appear to meet at a junction location. The Segmentation Map contains uniquely numbered disjoint regions bound by the contours. The Contour Map and Segmentation Maps are just a convenient way of representing the edge information. Only the locations at which exactly 3 distinct contours meet in the Contour Map (Figure 3 (B)) and correspondingly the locations at which exactly 3 distinct segments meet in the Segmentation Map (Figure 3 (C)) are considered for TJunction determination. Such locations can be easily determined from the Segmentation and Contour maps.
As shown in Figure 3E and 3F, at each junction location we have three regions, , and and contours, , and meeting. At each such junction, a circular mask of pixels is applied and the corresponding small patches of the segmentation map and contour map are used for further analysis. We determine the contours forming the “hat” of the TJunction (foreground) and the corresponding figure direction in two different ways: (1) based on the area of regions meeting at junction location within the small circular disk around junction; (2) based on the angle between contours meeting at the junction location. Finally, only those junctions locations for which figure direction, as determined based on both methods, is matching are introduced into the FGO model as TJunction local cues. Matching based on two different methods improves the overall accuracy in correctly identifying the “hat” (foreground) and “stem” (background) of TJunctions, in effect the correct figure direction.
The local neighborhood of TJunction influence is set to be a circular region of radius 15 pixels. All the border pixels near the junction location within a radius of 15 pixels that belong to the “hat” of the TJunction are set to +1 for the appropriate BO direction. Remember that for each orientation, we will have two TJunction maps, one for the BO preference direction, denoted as and the other for the opposite BO preference, denoted as . A pixel in is set to +1 if the direction of figure, as determined by both methods (Sections 4.2.1 and 4.2.1) is , i.e. “stem” of the TJunction is in the direction. Similarly, computed. The TJunction maps thus obtained are decomposed into multiscale pyramids, , where superscript, denotes scale, by successive downsampling, which are used to compute the cue specific BO pyramids as explained in Section 3.3, Equation 21.
4.2.1 Area based TJunction determination
Let , and be the three regions at a junction location (Figure 3E). After extracting the circular region around the junction by applying a circular mask of radius, 6 pixels, we count the number of pixels belonging to each of the regions, . The region, having the largest pixel count is determined as the figure region. In Figure 3E, is the region with largest pixel count, hence determined as the foreground. The contours abutting the figure region, as determined by pixel count, which are and (Figure 3F), form the “hat” of the TJunction. Contour forms the “stem” of the TJunction, which belongs to the background.
The local orientation at each contour location is known. Vectors of length pixels, normal to the local orientation are drawn at each “hat” contour location within the pixel neighborhood. If the normal vector intersects the figure region, , as determined based on region area, the edge/contour location is given a value of +1 in the TJunction map for the appropriate BO direction, which can be or . This is done for every pixel in the edge/contour map within a neighborhood of 15 pixel radius around the TJunction location for those contours ( and ) that form the “hat” of the TJunction. For example, in Figures 3E and 3F, if the local orientation of and is roughly 0, then the end point of normal vector in the direction intersects with the figure region, , as determined based on the segment area. So, the TJunction map for BO preference direction is set to +1 within the circular neighborhood of pixels. The TJunction map for BO direction will be zero.
4.2.2 Angle based TJunction determination
In this method, as in Section 4.2.1, a small circular patch of radius, 7 pixels is extracted from the contour map around Tjunction location. Pixels belonging to each contour, meeting at the junction are labeled with a distinct number, so for each contour, we track the first 7 pixels starting from the junction location. Since the starting point for each contour, is the same, the total angle at junction location is . For each contour, we define a vector (red arrows in Figure 3F) from the junction location to the last tracked point on the contour. We then compute the angle between the vectors corresponding to contours. The contours between which angle is the largest form the “hat” of the Tjunction. For example, in Figure 3F, is the angle between and , which is also the largest of the three angles, , and . So, in the angle based Tjunction computation also, and are determined to form the “hat” of the Tjunction. The figure direction at every pixel of the “hat” contours is determined as in Section 4.2.1.
Among all the potential TJunctions determined using the angle based method, potential YJunctions and Arrow junctions are discarded based on the angle formed by the contours at junction location. If the largest angle is greater than , such junctions are discarded. Since the largest angle greater than is typically seen in the case of Arrowjunctions, we do not include them in the computation. Arrow junctions appear in a scene when the corner of a 3D structure is seen from outside. In the same way, if each angle at a junction location is within , such junctions are discarded as those are most likely YJunctions. YJunctions appear in a scene when a 3D corner is viewed from inside the object, for example, corner of a room viewed from inside the room. Rest of the TJunctions are included in our computation. Angle based filtering of potential Arrow or YJunctions was not considered in previous methods Palou and Salembier (2011, 2013).
TJunctions and their figure directions are determined using both Segment Area based and Contour Angle based methods and the TJunctions are incorporated into the model only in those cases, where both methods give matching figure direction, which makes TJunction determination more accurate.
Accurately determining the figure side of a Tjunction from a small neighborhood of 67 pixel radius is quite challenging because, within that small neighborhood we generally do not have any information indicative of figure/ground relations, other than contour angle and segment area. Even though key point detection is a well studied area, hence locating a TJunction is not problematic, deciding which of the three regions is the foreground based on information from a small neighborhood is extremely challenging. So, when locally determining figure side of a Tjunction, segment area and contour angle were found to be the most exploitable properties.
5 Data and methods
The figureground dataset, a subset of BSDS 300 dataset, consists of 200 images of size pixels, where each image has two ground truth figureground labels Ren et al. (2006) and corresponding boundary maps. For each image, the two sets of figureground labels labels are annotated by users other than those who outlined the boundary maps. The figureground boundary consists of figure side of the boundary marked by +1 and the ground side boundary by 1.
The figureground classification accuracy (FGCA) for an image we report is the percentage of the total number of boundary pixels in the ground truth figure/ground label map for which a correct figure/ground classification decision is made by the model described in Section 3. Even though the model computes BO response at every location where cells are active, the BO responses are compared only at those locations for which ground truth figure/ground labels exist.
Whenever the two ground truth label maps differ for the same image, average of the FGCA for both ground truth label maps is reported. Since different figureground labelers interpret figure and ground sides differently depending on the context, such differences arise, as a result, the selfconsistency between figureground labelings between the two sets of ground truth annotations is 88%, which is the maximum achievable FGCA for the dataset. At each pixel, the direction of figure, as determined by the model can be correct or wrong. So, the average FGCA for the entire dataset, at chance is 50%, assuming figure/ground relations at neighboring pixels are independent. This assumption is consistent with previously reported results Ren et al. (2006), where same assumption was made. The complete details of the figureground dataset can be found in Martin et al. (2001); Ren et al. (2006); Fowlkes et al. (2007).
The entire BSDS figure/ground dataset consisting of 200 images is randomly split into training set of 100 images and test set of 100 images. Parameters of the model are tuned for the training dataset and the optimal values of parameters found for the training set are used to evaluate the FGCA of the test set of images. The average FGCA that we report for the entire test set is the average of FGCAs of all 100 images in the test set.
6 Results
To remind the readers, the model with only global cues of convexity, surroundedness and parallelism, without any local cues is referred to as the Reference model. As explained in Section 3, local cues, SA and TJunctions are added to the Orientation feature channel of the reference model. As we have previously described in Section 3.3, by setting and in Eq 22, the model with local cues can be reduced to the reference model. Similarly, by switching the weights for each local cue to zero, the effect of the other local cue on FGO can be studied. As explained in Section 3.3, the winning BO pyramids are upsampled to image resolution and summed across scales and feature channels (Eq 24) for each BO direction to get the response magnitude for that BO direction. The BO information derived this way is compared against the groundtruth from BSDS figure/ground dataset.
First, we wanted to quantify the performance of the reference model, which is devoid of both local cues, in terms of FGCA. With and , the overall FGCA for 100 test images was 58.44% (standard deviation = 0.1146). With only global cues, the 58.44% FGCA we achieved is 16.88% above chance level (50%). Hence, we can conclude that the global Gestalt properties of convexity, surroundedness and parallelism, which the reference model embodies, are important properties that are useful in FGO. The parameters used in the reference model computation are listed in Table 2. Unless stated otherwise explicitly, those parameters in Table 2 remain unchanged for the remaining set of results that we are going to discuss. Only the parameters specifically related to the addition of local cues are separately tuned and will be explicitly reported.
Parameter  Value 

0.5  
2.24  
1.57  
0.90  
2.70  
2.0  
1.0  
3.2  
0.8  
0.7854  
10 
Next, we wanted to study the effect of adding each local cue individually (Sections 6.1 and 6.2) and then the effect of both local cues together (Section 6.3).
6.1 Effect of adding Spectral Anisotropy
As explained in Section 4.1, Spectral Anisotropy was computed at the native resolution of the image by pooling Complex cell responses at many scales for each orientation. For each orientation, , two SA maps, and are created for respective antagonistic BO directions with respect to . The SA maps are then decomposed into multiscale pyramids by successively downsampling. The SA pyramids are then incorporated into the model as explained in Eq.20 and Eq.22. In this case, parameters and are tuned for the training dataset and is set to 0.
The parameter tuning procedure we use here is the same for other cases as well. We use multiresolution grid search for parameter tuning with the condition that the sum of tuned parameters should be 1. In this case, the condition was . We stop refining the resolution of the grid when the variation in FGCA upto second decimal point is zero, i.e. , only small changes are seen from third digit onward, after the decimal point.
The optimal parameters were found to be, and
for the training dataset. With these optimal parameter values, the FGCA for the test set was 62.69% (std. dev = 0.1204), which is a 7.3% improvement in the model’s performance after adding the local cue, Spectral Anisotropy, compared to the reference model’s FGCA of 58.44%. To verify if the improvement in FGCA that we see is statistically significant, we performed an unpaired sample, right tailed ttest (Table
3), where the null hypothesis was that the means of FGCAs of the reference model and the model with SA are equal. The alternate hypothesis was that the mean FGCA of the model with SA is higher than that of the reference model. The significance level,
was chosen. For other results (Sections 6.2, 6.3) as well, we do the same type of test, where the reference model’s FGCA is compared with that of modified model’s FGCA having different local cues. Hereafter, we refer to them as statistical tests.Statistical tests show that the mean FGCA of the model with SA is significantly higher than that of the reference model (). This demonstrates SA is a useful cue and can be successfully incorporated into the reference model, adding which results in statistically significant improvement in the model’s performance. This, and all other results are summarized in Table 3 for the test dataset.
6.2 Effect of adding TJunctions
As described in Section 4.2, TJunctions are computed at image resolution using the segmentation map and edge map obtained using the gPb Arbelaez et al. (2011) algorithm. Each of the TJunction maps for the 16 different BO directions is successively downsampled to create multiscale TJunction pyramids. The TJunction pyramids are incorporated into the model as explained in Eq.21 and Eq.22 and by setting . The other two parameters, and are tuned on the training dataset. With optimal parameter values, and (and ), the FGCA for the test set was found to be 59.48% (std. dev. = 0.1127). Compared to the reference model’s FGCA of 58.44%, we see that adding TJunctions improves the model’s performance in terms of FGCA by 1.78%. Based on the statistical tests (Table 3), we find that the improvement in FGCA that we see is indeed statistically significant.
6.3 Effect of adding both Spectral Anisotropy and TJunctions
SA is computed as explained in Section 4.1, TJunctions are computed as explained in Section 4.2, where TJunctions are derived from automatically extracted edges using the gPb algorithm. Both cues are added to the Reference model according to Eq 22. The parameters , and are tuned simultaneously on the training dataset using multiresolution grid search as before, with the constraint, . The optimal values of the parameters were found to be, , and . All other parameters remained unchanged as shown in Table 2. The FGCA of the combined model with both local cues, Spectral Anisotropy and TJunctions was 63.57% (std. dev = 0.1179) for the test dataset, which is higher that the FGCAs we obtained for the individual cues when they were added separately. We see an improvement in FGCA of 8.78% compared to that of the reference model with no local cues. As before, an unpaired sample, right tailed ttest comparing the reference model’s figure/ground decisions and the combined model’s figure/ground decisions with both SA and TJunctions showed statistically significant improvement (Table 3).
In addition to comparing the performance of the model with both local cues with the Reference model, we also compared the performance of the model with both local cues (Ref model + SA + TJunctions) to the model with only one (Ref model + SA) local cue. Unpaired sample righttailed ttests were used again with a significance level of 0.05. In this case the the null hypothesis is that adding TJunctions to the Reference Model with SA does not lead to statistically significant improvement in FGCA. The alternate hypothesis is that adding TJunctions leads to statistically significant improvement in FGCA when compared to the FGCA of Reference (global cues only) + SA model. Tests show adding TJunctions to the Reference + SA model leads to a statistically significant improvement (p = ).



pvalue  

Reference Model 

      
With SA 

7.3%  Yes  


1.78%  Yes  


8.78%  Yes  0 
In summary, we show that both SA and TJunctions are useful local cues of FGO, which produce statistically significant improvement in FGCA when added alone. When both cues are simultaneously present, they lead to even higher improvement in FGCA of the model indicating the cues are mutually facilitatory. An improvement of with only a few local and global cues at a minimal computational cost (see A for computational cost analysis) is truly impressive. Figures 4 and 5 show FGO results for some example images from the test dataset when both SA and TJunctions are added.
Next, we compare the performance of our model with state of the art methods for which all steps are fully automated (see Table 4). Here, we are comparing the FGCA of the model with both local cues with other methods that are not neurally inspired, instead are learning based and trained on thousands of images. Our model performs better than that of Maire (2010) even when it has only a few local and global FGO cues and not specifically tuned for best performance. The performance of our model is competitive with the state of the art models given our constraints discussed above, but leaves room for improvement. The performance of the model can be substantially improved by making several simple modifications and adding more local and global cues as discussed in Section 9.
Algorithm  FGCA 

M. Maire etal, ECCV, 2010 Maire (2010)  62% 
Our method  63.6% 
X. Ren etal, ECCV, 2006 Ren et al. (2006)  68.9% 
P. Salembier etal, IEEE TIP, 2013 Palou and Salembier (2013)  71.3% 
CL. Teo, etal, CVPR 2015 Teo et al. (2015)  74.7% 
D. Hoiem etal, ICCV 2007 Hoiem et al. (2011)  79% 
7 Discussion
We see improvement in FGCA of the combined model with both local cues. This improvement, from only two local cues added to one of the three feature channels is truly impressive. Moreover, the three feature channels (Color, Intensity and Orientation) were weighted equally. Better results can be achieved if we tune the weights for individual feature channels. But, since our objective here was to study how to integrate local and global cues and measure the relative importance of local cues in FGO, feature specific weight tuning was not done, but we consider to do this in future (See Section 9 for future work). Moreover, it is important to note that FGCA of the model with both local cues is always higher than the FGCAs of models with individual local cues. This suggests the local cues are mutually facilitatory, which is further validated by the fact that we see statistically significant improvement in FGCA when TJunctions are added as an additional cue to the Reference model having SA as one of the local cues.
We introduce a few novel methods in our work. First, demonstrating that Spectral Anisotropy can be computed with Simple and Complex cells found in area V1 is a novel contribution. The significance of this computation is that it demonstrates SA can be computed in low level visual areas, even in the striate cortex and it does not require specialized cells to detect these shading/texture gradients. Only a specific arrangement of Complex cells of various spatial frequencies on each side of the border is sufficient. These cues, first mathematically shown to be useful by Huggins et al. (2001), were psychophysically validated by Palmer and Ghose (2008). We showed these patterns are abundantly found in natural images Ramenahalli et al. (2011) and can be efficiently computed using 1D FFTs Ramenahalli et al. (2014). Now, we show that these cues can be computed in a biologically plausible manner, using Complex cells found commonly in striate cortex.
Next, in the detection of TJunctions, we filter out YJunctions and Arrow junctions using the angle property of these junction types. Since YJunctions and Arrow junctions are not occlusion cues, ideally those should not be considered as TJunctions, hence we device a method to remove such junctions. To the best of our knowledge, previous methods Palou and Salembier (2011, 2012, 2013) that use TJunctions as FGO cues have not looked closely at this issue, which we consider novel in our approach. Also, we explicitly compute the local figure/ground relations at a TJunction local based on local information, which is new. And, the way we organize the local cue computation such that the same computational routine can be used for incorporation of both cues into the model is noteworthy. With this, the implementation is made more efficient, allowing easy parallelization using Graphics Processing Units (GPU) and other hardware. Moreover, the combination of features and local cues is done at a late stage (Eq 22), which allows independent and parallel computation of features and local cues, which again makes the model computationally more efficient, allowing parallelization.
Even though we see FGCA improvement when Tjunctions are added, it is a relatively small, but statistically significant, improvement compared to that adding SA. Since TJunctions are generally regarded as strong cues of occlusion, this small, statistically significant improvement may seem counterintuitive. But it is important to note that TJunctions are extremely sparse, can be computed only at a few locations where exactly 3 different regions partially occlude each other, whereas SA can be computed at every border location of an object. Given the sparsity of Tjunctions, they can still be considered stronger FGO cues compared to SA. The presence of “inverted” TJunctions Palou and Salembier (2013, 2012, 2011), could also be the reason for diminished effect of TJunctions. From a computational cost perspective (A), even though the cost is , given their sparsity (typically 310 TJunctions per image), adding them as a local cue is justified.
Even though it is commonly assumed that TJunctions are unambiguous cues of occlusion, no systematic, datadriven analysis of the utility of TJunctions as a classic Gestalt cue was available until now. Moreover, there are few instances where researchers argue from the opposite perspective. Tse and Albert (1998) argue that high level surface and volume analysis takes place first, and only after such an analysis, a TJunction is interpreted to be an occlusion cue. As a result, we may not consciously notice the prevalence of “inverted” TJunctions.
The traditional view that TJunctions are unambiguous cues of occlusion has also been challenged by psychophysics experiments of McDermott (2004), where they find that making occlusion decisions from a small aperture, typically a few pixels wide, in real images is hard for humans. Some studies also suggest junctions in general, hence TJunctions, can be cues for image segmentation, but not for occlusion reasoning van der Helm (2011). These previous works and our own results do not support the generally held view that TJunctions are the most unambiguous occlusion cues. But, these cues are useful and produce statistically significant improvement in FGCA. This is an important contribution of our work.
While comparing the performance of our model with existing methods, as noted in Section 6.3, we need to keep in mind some important differences. First, our model is not trained on image features, hence generalization to any other dataset does not require additional training. Second, our model is neurally inspired, built to provide a general framework for incorporating and studying local and global Gestalt cues, not specifically optimized for best accuracy. Moreover, we use only a handful of cues, yet perform better than some existing models (Maire (2010) in Table 4). While Maire (2010) uses 64 different shapemes, descriptors of local shape derived from object edges, Ren et al. (2006) incorporates empirical frequencies of 4 different junction types derived from training data, in addition to shapemes in a Conditional Random Field based model. Also, Ren et al. (2006) compare figure/ground relations with the groundtruth only at a partial set of locations where their edge detection algorithm finds a matching edge with the groundtruth. It is not clear what percentage of edges match with the groundtruth. Palou and Salembier (2013) use 8 color based cues, in addition to TJunctions and local contour convexity in their model. The other two (Teo et al. (2015); Hoiem et al. (2011) ) models use a much larger number of cues to achieve FGO. Moreover, the models we are comparing with are neither strictly Gestalt cue based nor neurally motivated. To the best of our knowledge, there are no comparable neurally inspired, feedforward, fully automated models that are tested on the BSDS figureground dataset. The model proposed by Sakai et al. (2012) is tested on BSDS FG database, but it requires human drawn contours.
The method we use to report FGCA can be very different from the methods of other models listed in Table 4. We report the average of all pixels for all 100 test images, which can be considerably lower than computing the FGCA image by image and then averaging the FGCA of all images. It is not clear from other methods in Table 4, how the FGCA numbers were reported. Moreover, the exact split of the dataset into train and test set also has an effect. For some splits, the FGCA can be higher. The methods reported in Table 4 may not have used the same test/train split as it is not reported in previous methods. So, instead of comparing with existing methods in terms of absolute FGCA, a more appropriate way to look at our results would be from the perspective of relative improvement after adding each cue. From this perspective, we do see statistically significant improvement with the addition of each local cue. Moreover, our motivation in this study was always to quantify the utility of local and global cues and build a general framework to incorporate and study the effect of multiple local and global cues.
Lastly, we investigate if the influence of local cues should be strictly local or global. In our model, even though the local cues, SA and TJunctions, are computed based on the analysis of a strictly local neighborhood around the object boundary, they modulate the activity of cells at all scales, i.e. , their influence is global in nature. Should the influence of local cues be also local? To answer this question, we added local cues only at the top 2 layers of the model, tuned the optimal parameters, , and accordingly and recomputed FGCA. We found that with local cue influence at only the top two layers, the FGCA we obtained was lower than having them at all scales (See B for details). This confirms the influence of local cues should not be local, even though their computation should be strictly local to reduce the computational cost, which is the case in our model.
8 Conclusion
We develop a biologically motivated, feedforward computational model of FGO with local and global cues. Spectral Anisotropy and TJunctions are the local cues newly introduced into the model, which only influence the Orientation channel among the three feature channels. First, we show that even the reference model, with only a few global cues, convexity, surroundedness and parallelism, completely devoid of any local cues performs significantly better than chance level (50%) achieving a FGCA of 58.44% on the BSDS figureground dataset. Each local cue, when added alone leads to statistically significant improvement in the overall FGCA, compared to the reference model devoid of local cues, indicating their usefulness as independent local cues of FGO. The model with both SA and the TJunctions achieves an 8.77% improvement in terms of FGCA compared to that of the model without any local cues. Moreover, the FGCA of the model with both local cues is always higher than that of the models with individual local cues, indicating the mutually facilitatory nature of local cues. In conclusion, SA and TJunctions are useful, mutually beneficiary local cues and lead to statistically significant improvement in the FGCA of the feed forward, biologically motivated FGO model, either when added alone or together.
As we show in A, the computational complexity of adding both local cues is relatively low, yielding improvement in model’s performance. Given that the feature channel weights are unoptimized, model consists of only a few global and local cues, local cues added to only one of three feature channels and the model is not optimized for best FGCA^{2}^{2}2See Chapter 9 for a discussion on how FGCA of the model can be improved even with existing local cues., the performance of the model is highly impressive.
9 Future Work
In future, we intend to improve the FGCA of the model by tuning the inhibitory weight, for each feature and each local cue (Eqs. 18 – 21) and tuning feature specific weights in Eq 24. In addition, increasing the number of scales, having cells and cells of multiple radii can all lead to better FGCA. cells cells of multiple radii would capture the convexity and surroundedness cues better. Also, the model’s figureground response is computed by modulating the activity of cells, which are computed using Gabor filter kernels. The response of cells may not always exactly coincide with human drawn boundaries in the groundtruth, with which we compare the model’s response to calculate FGCA. Hence, averaging the BO response in a small pixel neighborhood and then comparing that with the groundtruth FG labels could yield improved FGCA. In future, we would like to explore these ideas in order to improve FGCA. Moreover, color based cues Troscianko et al. (1991); Zaidi and Li (2006), global cues such as symmetry Ardila et al. (2011) and medial axis Ardila et al. (2012) can be incorporated to improve the FGCA and make the model more robust.
In the biologically plausible SA computation 4.1, we used the Complex cell responses in all our computation. It would be interesting to see if similar or better FGCA can be achieved with Simple Even or Odd cells alone. In that case, the cost of computing SA would reduce by more than half. This would make the overall FGO model computation even more efficient. Also, in Section 4.1, filter size increment was in steps of 2 pixels. Having finer filter size resolutions (for example, instead of ) will be considered to improve the FGCA even more.
From a computational cost perspective, image segmentation using the gPb Arbelaez et al. (2011)
algorithm is the most expensive step in the FGO model with local cues. In order to decrease the computational cost, more efficient image segmentation algorithms should be explored. One efficient algorithm with similar performance as gPb (Fscore,
Arbelaez et al. (2011) = 0.70 vs. Fscore, Leordeanu et al. (2014) = 0.69 on BSDS 500 dataset) by Leordeanu et al. (2014) is a good candidate. Replacing gPb Arbelaez et al. (2011) algorithm with the algorithm by Leordeanu et al. (2014) for image segmentation, hence TJunction computation, can substantially reduce the computational overload, while achieving similar performance. Other recent methods with better image segmentation performance can also be considered. We would also consider parallelization of the model using GPUs and FPGAs in future.References
References
 Wagemans et al. (2012a) J. Wagemans, J. H. Elder, M. Kubovy, S. E. Palmer, M. A. Peterson, M. Singh, R. von der Heydt, A century of gestalt psychology in visual perception: I. perceptual grouping and figure–ground organization., Psychological bulletin 138 (2012a) 1172.
 Wagemans et al. (2012b) J. Wagemans, J. Feldman, S. Gepshtein, R. Kimchi, J. R. Pomerantz, P. A. van der Helm, C. van Leeuwen, A century of gestalt psychology in visual perception: II. conceptual and theoretical foundations., Psychological bulletin 138 (2012b) 1218.
 Koffka (1935) K. Koffka, Principles of Gestalt psychology, HarcourtBrace, New York, 1935.
 Bahnsen (1928) P. Bahnsen, Eine Untersuchung uber Symmetrie und Asymmetrie bei visuellen Wahrnehmungen, Zeitschrift fur Psychologie 108 (1928) 129–154.
 Palmer (1999) S. E. Palmer, Vision SciencePhotons to Phenomenology, MIT Press, Cambridge, MA, 1999.
 Fowlkes et al. (2007) C. Fowlkes, D. Martin, J. Malik, Local figureground cues are valid for natural images, Journal of Vision 7 (2007).
 Heitger et al. (1992) F. Heitger, L. Rosenthaler, R. von ver Heydt, E. Peterhans, O. Kübler, Simulation of neural contour mechanisms: from simple to endstopped cells, Vision Research 32 (1992) 963 – 981.

Huggins et al. (2001)
P. Huggins, H. Chen,
P. Belhumeur, S. Zucker,
Finding folds: On the appearance and identification
of occlusion,
in: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 2, IEEE, pp. II–718.
 Palmer and Ghose (2008) S. Palmer, T. Ghose, Extremal edges: A powerful cue to depth perception and figureground organization, Psychological Science 19 (2008) 77–84.
 Ramenahalli et al. (2011) S. Ramenahalli, S. Mihalas, E. Niebur, Extremal edges: Evidence in natural images, in: 45th Annual Conference on Information Sciences and Systems (CISS), 2011, pp. 1 –5.
 Zhou et al. (2000) H. Zhou, H. S. Friedman, R. von der Heydt, Coding of border ownership in monkey visual cortex, J. Neurosci. 20 (2000) 6594–6611.
 Williford and von der Heydt (2016) J. R. Williford, R. von der Heydt, Figureground organization in visual cortex for natural scenes, eNeuro 3 (2016).
 Craft et al. (2007) E. Craft, H. Schutze, E. Niebur, R. Von Der Heydt, A neural model of figureground organization, Journal of Neurophysiology 97 (2007) 4310–4326.
 Roelfsema et al. (2002) P. R. Roelfsema, V. A. Lamme, H. Spekreijse, H. Bosch, Figure ground segregation in a recurrent network architecture, Journal of Cognitive Neuroscience 14 (2002) 525–537.
 Zhaoping (2005) L. Zhaoping, Border ownership from intracortical interactions in visual area V2, Neuron 47 (2005) 143–153.
 Mihalas et al. (2011) S. Mihalas, Y. Dong, R. von der Heydt, E. Niebur, Mechanisms of perceptual organization provide autozoom and autolocalization for attention to objects, Proceedings of the National Academy of Sciences 108 (2011) 7583–8.
 Hu et al. (2019) B. Hu, R. von der Heydt, E. Niebur, Figureground organization in natural scenes: Performance of a recurrent neural model compared with neurons of area v2, eNeuro 6 (2019).
 Ramenahalli et al. (2014) S. Ramenahalli, S. Mihalas, E. Niebur, Local spectral anisotropy is a valid cue for figure–ground organization in natural scenes, Vision research 103 (2014) 116–126.
 Ramenahalli et al. (2012) S. Ramenahalli, S. Mihalas, E. Niebur, Figureground classification based on spectral anisotropy of local image patches, in: Proceedings of the 46th Annual IEEE Conference on Information Sciences and Systems (IEEECISS), 2012, pp. 1–5.
 Heitger and von der Heydt (1993) F. Heitger, R. von der Heydt, A computational model of neural contour processing: figureground segregation and illusory contours, in: Proc. 4th Int. Conf. Computer Vision, IEEE Computer Society Press, 1993, pp. 32–40.
 Hansen and Neumann (2002) T. Hansen, H. Neumann, A biologically motivated scheme for robust junction detection, Proceedings of Second International Workshop on Biologically Motivated Computer Vision (2002) 16–26.
 Rubin (1921) E. Rubin, Visuell wahrgenommene Figuren, Kobenhaven: Glydenalske Boghandel, 1921.
 Wertheimer (1923) M. Wertheimer, Untersuchungen zur Lehre von der Gestalt II, Psychol. Forsch. 4 (1923) 301–350.
 Lamme (1995) V. A. Lamme, The neurophysiology of figureground segregation in primary visual cortex, The Journal of Neuroscience 15 (1995) 1605–1615.
 Super and Lamme (2007) H. Super, V. A. Lamme, Altered figureground perception in monkeys with an extrastriate lesion, Neuropsychologia 45 (2007) 3329–3334.
 Williford and von der Heydt (2014) J. R. Williford, R. von der Heydt, Early visual cortex assigns border ownership in natural scenes according to image context, Journal of Vision 14 (2014) 588–588.
 Ren et al. (2006) X. Ren, C. C. Fowlkes, J. Malik, Figure/ground assignment in natural images, in: European Conference on Computer Vision, Springer, 2006, pp. 614–627.
 Hoiem et al. (2011) D. Hoiem, A. A. Efros, M. Hebert, Recovering occlusion boundaries from an image, International Journal of Computer Vision 91 (2011) 328–346.
 Teo et al. (2015) C. L. Teo, C. Fermüller, Y. Aloimonos, Fast 2D border ownership assignment, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2015) 5117–5125.
 Hoiem et al. (2007) D. Hoiem, A. N. Stein, A. A. Efros, M. Hebert, Recovering occlusion boundaries from a single image, in: IEEE 11th International Conference on Computer Vision, ICCV, 2007, pp. 1–8.
 Amer et al. (2010) M. R. Amer, R. Raich, S. Todorovic, Monocular extraction of 2.1D sketch, in: Proceedings of the International Conference on Image Processing, ICIP 2010, September 2629, Hong Kong, China, pp. 3437–3440.
 Amer et al. (2015) M. R. Amer, S. Yousefi, R. Raich, S. Todorovic, Monocular extraction of 2.1D sketch using constrained convex optimization, International Journal of Computer Vision 112 (2015) 23–42.
 Leichter and Lindenbaum (2009) I. Leichter, M. Lindenbaum, Boundary ownership by lifting to 2.1D, in: IEEE 12th International Conference on Computer Vision, 2009, IEEE, pp. 9–16.
 Palou and Salembier (2013) G. Palou, P. Salembier, Monocular depth ordering using Tjunctions and convexity occlusion cues., IEEE Transactions on Image Processing 22 (2013) 1926–1939.

Palou and Salembier (2012)
G. Palou, P. Salembier,
From local occlusion cues to global monocular depth estimation,
in: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2012, IEEE, pp. 793–796.  Palou and Salembier (2011) G. Palou, P. Salembier, Occlusionbased depth ordering on monocular images with binary partition tree, in: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2011, IEEE, pp. 1093–1096.
 Salembier and Garrido (2000) P. Salembier, L. Garrido, Binary partition tree as an efficient representation for image processing, segmentation, and information retrieval, IEEE Transactions on Image Processing 9 (2000) 561–576.
 Nishigaki et al. (2012) M. Nishigaki, C. Fermüller, D. DeMenthon, The image torque operator: A new tool for midlevel vision, in: IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, pp. 502–509.
 Yu et al. (2001) S. X. Yu, T. S. Lee, T. Kanade, A hierarchical markov random field model for figureground segregation, Proceedings of Third International Workshop on Energy Minimization Methods in Computer Vision and Pattern Recognition (2001) 118–133.
 Baek and Sajda (2005) K. Baek, P. Sajda, Inferring figureground using a recurrent integrateandfire neural circuit, IEEE Transactions on Neural Systems and Rehabilitation Engineering 13 (2005) 125–130.
 Maire (2010) M. Maire, Simultaneous segmentation and figure/ground organization using angular embedding, in: European Conference on Computer Vision–ECCV, Springer, 2010, pp. 450–464.
 Yu (2009) S. Yu, Angular embedding: from jarring intensity differences to perceived luminance, in: IEEE Conference on Computer Vision and Pattern Recognition, 2009., IEEE, pp. 2302–2309.
 Ion et al. (2011) A. Ion, J. Carreira, C. Sminchisescu, Image segmentation by figureground composition into maximal cliques, in: IEEE International Conference on Computer Vision, IEEE, pp. 2110–2117.
 Ion et al. (2014) A. Ion, J. Carreira, C. Sminchisescu, Probabilistic joint image segmentation and labeling by figureground composition, International Journal of Computer Vision 107 (2014) 40–57.
 Kogo et al. (2010) N. Kogo, C. Strecha, L. Van Gool, J. Wagemans, Surface construction by a 2D differentiation–integration process: A neurocomputational model for perceived border ownership, depth, and lightness in kanizsa figures., Psychological review 117 (2010) 406.
 Froyen et al. (2010) V. Froyen, J. Feldman, M. Singh, A bayesian framework for figureground interpretation, in: J. Lafferty, C. Williams, J. ShaweTaylor, R. Zemel, A. Culotta (Eds.), Advances in Neural Information Processing Systems 23, Curran Associates, Inc., 2010, pp. 631–639.
 Kienker et al. (1986) P. K. Kienker, T. J. Sejnowski, G. E. Hinton, L. E. Schumacher, Separating figure from ground with a parallel network, Perception 15 (1986) 197–216.
 Grossberg and Mingolla (1985) S. Grossberg, E. Mingolla, Neural dynamics of form perception: boundary completion, illusory figures, and neon color spreading., Psychological review 92 (1985) 173.
 Grossberg (1994) S. Grossberg, 3D vision and figureground separation by visual cortex, Perception & psychophysics 55 (1994) 48–121.
 Sajda and Finkel (1995) P. Sajda, L. Finkel, Intermediatelevel visual representations and the construction of surface perception, J Cogn Neurosci 7 (1995) 267–291.
 Jehee et al. (2007) J. F. Jehee, V. A. Lamme, P. R. Roelfsema, Boundary assignment in a recurrent network architecture, Vision research 47 (2007) 1153–1165.
 Li (2003) Z. Li, V1 mechanisms and some figure–ground and border effects, Journal of PhysiologyParis 97 (2003) 503–515.
 Li (2000) Z. Li, Can V1 mechanisms account for figureground and medial axis effects?, in: S. A. Solla, T. K. Leen, K. Müller (Eds.), Advances in Neural Information Processing Systems 12, MIT Press, 2000, pp. 136–142.
 Kapadia et al. (1995) M. K. Kapadia, M. Ito, C. D. Gilbert, G. Westheimer, Improvement in visual sensitivity by changes in local context: Parallel studies in human observers and in V1 of alert monkeys, Neuron 15 (1995) 843 – 856.
 Slllito et al. (1995) A. M. Slllito, K. L. Grieve, H. E. Jones, J. Cudeiro, J. Davls, Visual cortical mechanisms detecting focal orientation discontinuities, Nature 378 (1995) 492–496.
 Knierim and Van Essen (1992) J. J. Knierim, D. C. Van Essen, Neuronal responses to static texture patterns in area V1 of the alert macaque monkey, J. Neurophysiology 67 (1992) 961–980.

Kikuchi and Akashi (2001)
M. Kikuchi, Y. Akashi,
A model of borderownership coding in early vision,
in: International Conference on Artificial Neural Networks – ICANN, Springer, 2001, pp. 1069–1074.
 Russell et al. (2014) A. F. Russell, S. Mihalas, R. von der Heydt, E. Niebur, R. EtienneCummings, A model of protoobject based saliency, Vision Research 94 (2014) 1–15.
 Molin et al. (2013) J. L. Molin, A. F. Russell, S. Mihalas, E. Niebur, R. EtienneCummings, Protoobject based visual saliency model with a motionsensitive channel, in: Biomedical Circuits and Systems Conference (BioCAS), 2013 IEEE, pp. 25–28.
 Hu and Niebur (2017) B. Hu, E. Niebur, A recurrent neural model for protoobject based contour integration and figureground segregation, Journal of Computational Neuroscience (2017).
 Layton et al. (2012) O. W. Layton, E. Mingolla, A. Yazdanbakhsh, Dynamic coding of borderownership in visual cortex, Journal of vision 12 (2012) 8.
 Domijan and Šetić (2008) D. Domijan, M. Šetić, A feedback model of figureground assignment, Journal of vision 8 (2008) 10.
 Sakai et al. (2012) K. Sakai, H. Nishimura, R. Shimizu, K. Kondo, Consistent and robust determination of border ownership based on asymmetric surrounding contrast, Neural Networks 33 (2012) 257–274.
 Nishimura and Sakai (2004) H. Nishimura, K. Sakai, Determination of border ownership based on the surround context of contrast, Neurocomputing 58 (2004) 843–848.
 Nishimura and Sakai (2005) H. Nishimura, K. Sakai, The computational model for borderownership determination consisting of surrounding suppression and facilitation in early vision, Neurocomputing 65 (2005) 77–83.
 Russell et al. (2014) A. F. Russell, S. Mihalaş, R. von der Heydt, E. Niebur, R. EtienneCummings, A model of protoobject based saliency, Vision research 94 (2014) 1–15.
 Rensink (2000) R. A. Rensink, The dynamic representation of scenes, Visual cognition 7 (2000) 17–42.
 Adelson and Bergen (1985) E. Adelson, J. Bergen, Spatiotemporal energy models for the perception of motion, J. Opt. Soc. Am. A. 2 (1985) 284–299.
 MATLAB (9 30) MATLAB, 2D crosscorrelation, https://www.mathworks.com/help/signal/ref/xcorr2.html, Accessed: 20130930.
 Itti et al. (1998) L. Itti, C. Koch, E. Niebur, A model of saliencybased visual attention for rapid scene analysis, IEEE Transactions on Pattern Analysis and Machine Intelligence 20 (1998) 1254 –1259.
 Weisstein (9 30) E. W. Weisstein, von Mises Distribution, http://mathworld.wolfram.com/vonMisesDistribution.html, Accessed: 20140930.
 Arbelaez et al. (2011) P. Arbelaez, M. Maire, C. Fowlkes, J. Malik, Contour detection and hierarchical image segmentation, IEEE Trans. Pattern Anal. Mach. Intell. 33 (2011) 898–916.
 Martin et al. (2001) D. Martin, C. Fowlkes, D. Tal, J. Malik, A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics, in: Proceedings of Eighth IEEE International Conference on Computer Vision, 2001, volume 2, IEEE, pp. 416–423.
 Palmer and Ghose (2008) S. Palmer, T. Ghose, Extremal Edge– A Powerful Cue to Depth Perception and FigureGround Organization, Psychological Science 19 (2008) 77.
 Tse and Albert (1998) P. U. Tse, M. K. Albert, Amodal completion in the absence of image tangent discontinuities, Perception 27 (1998) 455–464.
 McDermott (2004) J. McDermott, Psychophysics with junctions in real images, Perception 33 (2004) 1101–1127.
 van der Helm (2011) P. A. van der Helm, Bayesian confusions surrounding simplicity and likelihood in perceptual organization, Acta psychologica 138 (2011) 337–346.
 Troscianko et al. (1991) T. Troscianko, R. Montagnon, J. L. Clerc, E. Malbert, P.L. Chanteau, The role of colour as a monocular depth cue, Vision Research 31 (1991) 1923 – 1929.
 Zaidi and Li (2006) Q. Zaidi, A. Li, Threedimensional shape perception from chromatic orientation flows, Visual Neuroscience 23 (2006) 323–330.
 Ardila et al. (2011) D. Ardila, S. Mihalas, E. Niebur, How perceptual grouping affects the salience of symmetry, in: Annual Meeting, Society for Neuroscience, Washington DC, p. Abstract 801.01/LL18.
 Ardila et al. (2012) D. Ardila, S. Mihalas, R. von der Heydt, E. Niebur, Medial axis generation in a model of perceptual organization, 46th IEEE Annual Conference on Information Sciences and Systems (2012) 1–4.
 Leordeanu et al. (2014) M. Leordeanu, R. Sukthankar, C. Sminchisescu, Generalized boundaries from multiple image interpretations, IEEE transactions on pattern analysis and machine intelligence 36 (2014) 1312–1324.
Appendix A Computational complexity of adding local cues
The most computationally intensive part of SA computation is the correlations involved in Eq 25, which has a computational complexity of when implemented in Fourier domain, where and are the number of rows and columns in the image.
The computationally intensive part of Tjunction computation is the gPb Arbelaez et al. (2011) based image segmentation. We utilize this algorithm as is, hence we will not delve into exact estimation of computational complexity for this step. Once the contours and segmentation maps are obtained using gPb algorithm, the computation of each TJunction using both methods described in Section 4.2.1 and Section 4.2.2 involves multiplying the edge maps, segmentation maps with masks of appropriate sizes, counting and tracking pixels, computing angles, etc, which roughly translates into a computational complexity of for both methods, where pixels for Segment Area based TJunction computation (Section 4.2.1) and pixels for Contour Angle based TJunction computation (Section 4.2.2). Typically
TJunctions are found in an image. So, once edges/segmentation map is computed, since only few TJunctions are typically present in images and the size of mask is not very large, subsequent computation is not very time consuming. With appropriate modifications, it should be possible to reduce the computational complexity of TJunction determination even further, which is not optimized at the moment.
Appendix B Local cues influencing only top 2 layers
Should the influence of local cues also be strictly local? Local cues, by definition, should be computed based on the analysis of a small patch of an image to determine figureground relations. This is what makes them computationally more efficient. But, should their influence also be local? There is no a priori reason why their influence should be strictly local. To verify whether there is higher benefit in adding them locally only at the top layer (i.e. , at native image resolution only), we added them only at the top layer. For SA it resulted in a noticeable, but very small improvement. For TJunctions, the change was barely noticeable. This could be due to extremely small size of von Mises filter kernels that we use ( pixels) in comparison with the images size ( pixels). So, we added the local cues to the top two layers. For each local cue added separately, the optimal parameters of the model were recomputed and those parameters were used to compute the FGCA. The versions of the model with local cues only at the top 2 layers did not give rise to better FGCA than what we saw earlier with the cues added at all scales. The results are summarized in Table 5.
Model  k = 2  k = 10  

Ref Model    58.44%  
Ref + SA  62.42%  62.69%  

59.12%  59.48% 
Comments
There are no comments yet.