Limits...
Extraction of surface-related features in a recurrent model of V1-V2 interactions.

Weidenbacher U, Neumann H - PLoS ONE (2009)

Bottom Line: The approach is based on feedforward and feedback mechanisms found in visual cortical areas V1 and V2.Unlike previous proposals which treat localized junction configurations as 2D image features, we link them to mechanisms of apparent surface segregation.As a consequence, we demonstrate how junctions can change their perceptual representation depending on the scene context and the spatial configuration of boundary fragments.

View Article: PubMed Central - PubMed

Affiliation: Institute of Neural Information Processing, University of Ulm, Ulm, Germany. ulrich.weidenbacher@uni-ulm.de

ABSTRACT

Background: Humans can effortlessly segment surfaces and objects from two-dimensional (2D) images that are projections of the 3D world. The projection from 3D to 2D leads partially to occlusions of surfaces depending on their position in depth and on viewpoint. One way for the human visual system to infer monocular depth cues could be to extract and interpret occlusions. It has been suggested that the perception of contour junctions, in particular T-junctions, may be used as cue for occlusion of opaque surfaces. Furthermore, X-junctions could be used to signal occlusion of transparent surfaces.

Methodology/principal findings: In this contribution, we propose a neural model that suggests how surface-related cues for occlusion can be extracted from a 2D luminance image. The approach is based on feedforward and feedback mechanisms found in visual cortical areas V1 and V2. In a first step, contours are completed over time by generating groupings of like-oriented contrasts. Few iterations of feedforward and feedback processing lead to a stable representation of completed contours and at the same time to a suppression of image noise. In a second step, contour junctions are localized and read out from the distributed representation of boundary groupings. Moreover, surface-related junctions are made explicit such that they are evaluated to interact as to generate surface-segmentations in static images. In addition, we compare our extracted junction signals with a standard computer vision approach for junction detection to demonstrate that our approach outperforms simple feedforward computation-based approaches.

Conclusions/significance: A model is proposed that uses feedforward and feedback mechanisms to combine contextually relevant features in order to generate consistent boundary groupings of surfaces. Perceptually important junction configurations are robustly extracted from neural representations to signal cues for occlusion and transparency. Unlike previous proposals which treat localized junction configurations as 2D image features, we link them to mechanisms of apparent surface segregation. As a consequence, we demonstrate how junctions can change their perceptual representation depending on the scene context and the spatial configuration of boundary fragments.

Show MeSH

Related in: MedlinePlus

Evaluation of extracted junction signals on synthetic test image.A synthetic test image (A) reproduced from Smith and Brady (1997) was used to evaluate and compare extracted junction signals against a computational scheme (structure tensor) for corner detection proposed by Harris and Stephens (1988). The extracted junction saliency is visualized in a feature likelihood map (B) and detected junction positions/types are superimposed on input image (C). ROC curves are computed from structure tensor results (dashed), initial model responses (dotted) and from converged model responses after 4 recurrent cycles (solid) (D). Abscissa denotes the false alarm rate, and the ordinate denotes the hit rate. Note, that for better visibility the abscissa has been scaled to [0, 0.1].
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2691604&req=5

pone-0005909-g014: Evaluation of extracted junction signals on synthetic test image.A synthetic test image (A) reproduced from Smith and Brady (1997) was used to evaluate and compare extracted junction signals against a computational scheme (structure tensor) for corner detection proposed by Harris and Stephens (1988). The extracted junction saliency is visualized in a feature likelihood map (B) and detected junction positions/types are superimposed on input image (C). ROC curves are computed from structure tensor results (dashed), initial model responses (dotted) and from converged model responses after 4 recurrent cycles (solid) (D). Abscissa denotes the false alarm rate, and the ordinate denotes the hit rate. Note, that for better visibility the abscissa has been scaled to [0, 0.1].

Mentions: In this section, we evaluate our model by comparing our recurrent junction detection scheme with results obtained by simply switching the recurrent feedback cycle off, which reduces the model to an ordinary feedforward model. Moreover, we compare our results to a standard computer vision corner detection scheme based on the Harris corner detector [4]. In a comparative study of different corner detection schemes, the Harris corner detector provides the best results among five corner detectors [42]. As input for our comparison, we use corner test image adapted from Smith and Brady [5] that poses several challenges such as, e.g., low contrast regions, smooth luminance gradients, or obtuse and acute angles. Moreover, all types of junctions (L, T and X) considered are represented in the test image together with information about their exact position (ground truth information). Since the Harris corner detector is not able to discriminate between different junction types, our comparison is only based on the detection performance, irrespective of the junction category. To measure the performance of the different schemes we use receiver operator characteristic (ROC) curves. This method is frequently used to evaluate true positive rate or hit rate and the false positive rate of a binary classifier system as its discrimination threshold is varied. Here, we use the junction feature map as input for the ROC analysis. Figure 14 shows the resulting ROC curves extracted from junction feature map given the test image as input. It is clearly visible that the ROC curve computed from the recurrent model responses lies well above the Harris corner detector curve and the initial feedforeward model curve. This suggest for a significantly better detection performance of our recurrent model compared to feedforward processing-based junction detection schemes.


Extraction of surface-related features in a recurrent model of V1-V2 interactions.

Weidenbacher U, Neumann H - PLoS ONE (2009)

Evaluation of extracted junction signals on synthetic test image.A synthetic test image (A) reproduced from Smith and Brady (1997) was used to evaluate and compare extracted junction signals against a computational scheme (structure tensor) for corner detection proposed by Harris and Stephens (1988). The extracted junction saliency is visualized in a feature likelihood map (B) and detected junction positions/types are superimposed on input image (C). ROC curves are computed from structure tensor results (dashed), initial model responses (dotted) and from converged model responses after 4 recurrent cycles (solid) (D). Abscissa denotes the false alarm rate, and the ordinate denotes the hit rate. Note, that for better visibility the abscissa has been scaled to [0, 0.1].
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2691604&req=5

pone-0005909-g014: Evaluation of extracted junction signals on synthetic test image.A synthetic test image (A) reproduced from Smith and Brady (1997) was used to evaluate and compare extracted junction signals against a computational scheme (structure tensor) for corner detection proposed by Harris and Stephens (1988). The extracted junction saliency is visualized in a feature likelihood map (B) and detected junction positions/types are superimposed on input image (C). ROC curves are computed from structure tensor results (dashed), initial model responses (dotted) and from converged model responses after 4 recurrent cycles (solid) (D). Abscissa denotes the false alarm rate, and the ordinate denotes the hit rate. Note, that for better visibility the abscissa has been scaled to [0, 0.1].
Mentions: In this section, we evaluate our model by comparing our recurrent junction detection scheme with results obtained by simply switching the recurrent feedback cycle off, which reduces the model to an ordinary feedforward model. Moreover, we compare our results to a standard computer vision corner detection scheme based on the Harris corner detector [4]. In a comparative study of different corner detection schemes, the Harris corner detector provides the best results among five corner detectors [42]. As input for our comparison, we use corner test image adapted from Smith and Brady [5] that poses several challenges such as, e.g., low contrast regions, smooth luminance gradients, or obtuse and acute angles. Moreover, all types of junctions (L, T and X) considered are represented in the test image together with information about their exact position (ground truth information). Since the Harris corner detector is not able to discriminate between different junction types, our comparison is only based on the detection performance, irrespective of the junction category. To measure the performance of the different schemes we use receiver operator characteristic (ROC) curves. This method is frequently used to evaluate true positive rate or hit rate and the false positive rate of a binary classifier system as its discrimination threshold is varied. Here, we use the junction feature map as input for the ROC analysis. Figure 14 shows the resulting ROC curves extracted from junction feature map given the test image as input. It is clearly visible that the ROC curve computed from the recurrent model responses lies well above the Harris corner detector curve and the initial feedforeward model curve. This suggest for a significantly better detection performance of our recurrent model compared to feedforward processing-based junction detection schemes.

Bottom Line: The approach is based on feedforward and feedback mechanisms found in visual cortical areas V1 and V2.Unlike previous proposals which treat localized junction configurations as 2D image features, we link them to mechanisms of apparent surface segregation.As a consequence, we demonstrate how junctions can change their perceptual representation depending on the scene context and the spatial configuration of boundary fragments.

View Article: PubMed Central - PubMed

Affiliation: Institute of Neural Information Processing, University of Ulm, Ulm, Germany. ulrich.weidenbacher@uni-ulm.de

ABSTRACT

Background: Humans can effortlessly segment surfaces and objects from two-dimensional (2D) images that are projections of the 3D world. The projection from 3D to 2D leads partially to occlusions of surfaces depending on their position in depth and on viewpoint. One way for the human visual system to infer monocular depth cues could be to extract and interpret occlusions. It has been suggested that the perception of contour junctions, in particular T-junctions, may be used as cue for occlusion of opaque surfaces. Furthermore, X-junctions could be used to signal occlusion of transparent surfaces.

Methodology/principal findings: In this contribution, we propose a neural model that suggests how surface-related cues for occlusion can be extracted from a 2D luminance image. The approach is based on feedforward and feedback mechanisms found in visual cortical areas V1 and V2. In a first step, contours are completed over time by generating groupings of like-oriented contrasts. Few iterations of feedforward and feedback processing lead to a stable representation of completed contours and at the same time to a suppression of image noise. In a second step, contour junctions are localized and read out from the distributed representation of boundary groupings. Moreover, surface-related junctions are made explicit such that they are evaluated to interact as to generate surface-segmentations in static images. In addition, we compare our extracted junction signals with a standard computer vision approach for junction detection to demonstrate that our approach outperforms simple feedforward computation-based approaches.

Conclusions/significance: A model is proposed that uses feedforward and feedback mechanisms to combine contextually relevant features in order to generate consistent boundary groupings of surfaces. Perceptually important junction configurations are robustly extracted from neural representations to signal cues for occlusion and transparency. Unlike previous proposals which treat localized junction configurations as 2D image features, we link them to mechanisms of apparent surface segregation. As a consequence, we demonstrate how junctions can change their perceptual representation depending on the scene context and the spatial configuration of boundary fragments.

Show MeSH
Related in: MedlinePlus