Computer and Robot vision

Assessing an image processing model of edge visibility with a psycho-visual experiment

Karine Joulan, Roland Brémond and Nicolas Hautière

An image processing simulation of the human visibility could be a valuable tool for a large spectrum of applications. We have developed a computational model which computes the visibility of edges in an image for a human observer [Joulan et al., 2011, CVPR workshop on Biologically-Consistent Vision, Colorado Springs, 33-38]. From this model, we compute the visibility as the ratio between contrast and contrast detection threshold, on the objects'edges. To assess the model predictions with respect to human performance, a psycho-visual experiment was performed on a detection task with 30 subjects. Four stimuli (discs and annulus with 1° and 0.2° in diameter) of various contrasts were displayed during 200 ms. The 50% detection rate was computed by fitting the experimental data with a Weibull function, and compared to the model's predictions. We found a good quantitative agreement for the discs, while the model underestimates the rings'visibility. These results suggest a new model, computing the object's visibility from the accumulation of local visibility indexes, rather than from their maximum.

The role of nonlinearities in hierarchical feed-forward models for pattern recognition

Sven Eberhardt, Tobias Kluth, Manfred Fahle and Christoph Zetzsche

Hierarchical feed-forward models for pattern recognition have been proven to be plausible in explaining receptive field properties of neurons in the ventral stream of primate visual cortex. In particular, the HMAX-model presented by [Serre et al, 2007, PNAS, 104, 6424-6429] can account for psychophysical results in rapid object categorization experiments by using alternations of matching and pooling layers. However, it is still unclear which nonlinearities are required in layers to perform pattern recognition tasks. Here, we developed a generalized form of the HMAX model in which we systemically varied the nonlinearities at each layer in the hierarchy between specificity and invariance using a control parameter. We evaluate the resulting model on several classification tasks and show that classification performance depends on the choice of control parameters for individual layers. Second, during a learning phase, we fed videos of natural scenes into the model and used both principal component analysis and slow feature analysis to extract components from feature space at different levels of the processing hierarchy. We show how the control parameters affect the extracted components and draw conclusions on how layers need to be arranged in an optimal feed-forward architecture for pattern recognition in natural scenes.

An a-contrario model for predicting perceptual thresholds in contours detection.

Samy Blusseau, José Lezama, Rafael Grompone and Jean-Michel Morel

This communication bridges a gap between computer vision methods and classic psychophysical setups. The non-accidentalness principle is invoked in psychophysics to explain the emergence of percepts. The mathematical theory of a contrario detection formalizes this principle [1] and attempts to predict ideal perception thresholds. Thus, it is natural to reconsider from a computational perspective, classic and new psychophysical experiments evaluating the human perception performance. To this aim, we chose the psychophysical experiments by Wagemans et al. [2] where subjects are presented with Gabor-rendered outlines of real world objects. In these experiments, orientation jitter was added to the elements with the aim of determining its effect on human object detection performance. Using the a contrario theory, the human detection thresholds can be compared rationally to the algorithmic ones. To allow a broader experimentation, we built an online web facility where users can perform object detection experiments, and compare their detection curves to the ones predicted analytically by the computational model. [1] Desolneux, A., Moisan, L., Morel, JM., 2008, From Gestalt Theory to Image Analysis, a Probabilistic Approach, Springer. [2] Nygård, G. E., Van Looy, T., & Wagemans, J., 2009, Vision Research, 49 (20), 2475-2484.

A motion feature - based algorithm for the detection of specular objects in natural scenes

Katja Doerschner and Ozgur Yilmaz

Successful identification of specularities in an image can be crucial for an artificial vision system when extracting the semantic content of an image or while interacting with the environment. We developed an algorithm that relies on scale and rotation invariant feature extraction techniques and uses motion cues to detect and localize specular surfaces in an office scene. Appearance change in feature vectors is used to quantify the appearance distortion on specular surfaces, which has previously been shown to be a powerful indicator for shininess [Doerschner et al, 2011, Current Biology 21(23), 2010-2016]. The algorithm combines epipolar deviations [Swaminathan et al, 2002, ECCV, 167-172] and appearance distortion cues and succeeds in localizing specular objects across a wide range of camera motions and speeds, object sizes and shapes, and performs well under image noise and blur conditions.