The analog implementation of a phase-based technique for disparity estimation is discussed. This technique is based on the convolution of images with Gabor filters. The article shows that by replacing the Gaussian envelope with other envelopes, the convolution operation is equivalent to the solution of a system of differential equations, whose order is related to the smoothness of the kernel. A detailed comparison between the disparity estimates obtained using these kernels and those obtained using the standard filter is presented. The discretization of the model leads to lattice networks in which the number of connections per node required to perform convolution is limited to the first few nearest neighbors. The short connection length makes these filter suitable for analog VLSI implementation, for which the number of connection per node is a crucial factor. Experimental measures on a prototype CMOS 17-node chip validated the approach.
We present an innovative framework for computer vision based on the principles of communication across heterogeneous visual modules and data fusion. Its goal is to support the building of advanced computer vision systems based on the principle of the diversification of the information sources. We show how it is possible to design a simple and efficient communication language that it is still general enough to support the necessities of the great majority of different visual modules and how, on the basis of the developed communication language, it is possible to devise an optimal communication scheme able to merge information from heterogeneous visual modules in an provably optimal way, solving the problem of merging cross-correlated information.
We propose a recursive post-processing algorithm to improve feature-maps, like disparity- or motion-maps, computed by early vision modules. The statistical distribution of the features is computed from the original feature-map and from this the most likely candidate for a correct feature is determined for every pixel. This process is performed automatically by a clustering algorithm which determines the feature candidates as the cluster centers in the distribution. After determining the feature candidates a cost function is computed for every pixel and a candidate will only replace the original feature if the cost is reduced. In this way a new feature-map is generated which, in the next iteration, serves as the basis for the computation of the updated feature distribution. Iterations are stopped if the total cost reduction is less than a pre-defined threshold. In general, our technique is able to reduce two of the most common problems that affect feature-maps, the sparseness, i.e., the presence of areas where the algorithm is not able to give meaningful measurements, and the blur. In order to show the efficacy of our approach, we apply the reclustering algorithm to several examples of increasing complexity, showing results for synthetic and natural images.
Stereoscopic depth analysis by means of disparity estimation has been a classical topic of computer vision, from the biological models of stereopsis up to the widely used techniques based on correlation or sum of squared differences. Most of the recent work on this topic has been devoted to the phase-based techniques, developed because of their superior performance and better theoretical grounding. In this article we characterize the performance of phase-based disparity estimators, giving quantitative measures of their precision and their limits, and how changes in contrast, imbalance, and noise in the two stereo images modify the attainable accuracy. We find that the theoretical range of measurable disparities, one period of the modulation of the filter, is not attainable: the actual range is approx. two-thirds of this value. We show that the phase-based disparity estimators are robust to changes in contrast of 100% or more and well tolerate imbalances of luminosity of 400% between the images composing the stereo pair. Clearing the Gabor filter of its DC component has been often advocated as a means to improve the accuracy of the results. We give a quantitative measure of this improvement and we show that using a DC-free Gabor filter leads to disparity estimators nearly insensitive to contrast and imbalance. Our tests show that the most critical source of error is noise: the error increases linearly with the increase of the noise level. We conclude studying the influence of the spectra and the luminosity of the input images on the error surface, both for artificial and natural images, showing that the spectral structure of the images has little influence on the results, changing only the form of the error surface near the limits of the detectable disparity range. In conclusion, this study allows estimation of the expected accuracy of custom-designed phase-based stereo analyzers for a combination of the most common error sources.
While optical flow has been often proposed for guiding a moving robot, its computational complexity has mostly prevented its actual use in real applications. We describe a restricted form of optical flow algorithm, which can be parallelized on chain-like neuronal structures, combining simplicity and speed. In addition, this algorithm makes use of predicted motion trajectories in order to remove noise from the input images.