Navigation auf uzh.ch
Navigation auf uzh.ch
Disclaimer: This research is not based on deep networks and has more of computer vision than machine learning.
2D Gabor filters are shown to model the early visual systems of mammals. Based on the filter responses of several complex-valued Gabor wavelets, so-called Gabor jets can be extracted as local texture descriptors around a certain offset point in an image. Specialized Gabor jet similarity functions allow to use these local descriptors for both the exact localization of textures in an image (Example) and the comparison or classification of those features.
One particular use case of Gabor jets developed in Statistical Gabor Graph Based Techniques for the Detection, Recognition, Classification, and Visualization of Human Faces (PDF, 11 MB) is to localize features from one image (or set of images) in another image. In a certain range around the correct offset position, this can even be done without a scanning process that other localization tools require, but two Gabor jets can estimate their respective disparity in a closed-form solution. This can be used to compute optical flow fields, for example, for estimating depth in stereo vision or tracking features in videos.
The task in this Bachelor's thesis is to investigate the potential of Gabor jet disparities in estimating optical flow and compare the results to other state-of-the-art techniques including deep learning methods. For example, stereo vision benchmark datasets such as ETH Stereo Benchmark can be used for evaluation.
