Fast and Robust Dense Correspondence Estimation

Description

This research project proposes to develop novel fast and robust dense correspondence algorithms. Dense correspondence estimation is a very fundamental task in computer vision, and has found widespread use in many tasks like stereo matching, optical flow estimation, video tracking, visual recognition, and camera calibration. However, the traditional correspondence algorithms are not robust to the changes of lighting, camera and exposure. Their performance is low around occlusions and low-texture regions. The computational complexity is high as well.We propose to investigate the dense correspondence problem along the following directions:- Linear-time illumination invariant dense correspondence estimation.We propose to develop efficient dense correspondence algorithm that is robust to both local and global affine illumination changes. We will formulate the computation of the affine transform between every two local image patches as an image denoising problem that can be solved efficiently and effectively using image filtering techniques. The transformed image intensity values can then be integrated with existing similarity measures for dense correspondence estimation.- Occlusion-aware dense correspondence estimation.We propose to develop novel efficient and robust edge-aware image filtering techniques and integrate them in the proposed correspondence algorithms for occlusions handling. A tree structure from the reference image will be extracted and used as the guidance for filtering. The resulted filter will be automatically edge-aware and can effectively reduce the incorrect correspondence estimates around occlusions.- Quantitative evaluation of the dense correspondence algorithms.A number of benchmarks with ground-truth correspondence obtained from either measurements of different active sensors or synthetic motions have been developed for specific computer vision tasks like stereo matching and optical flow estimation. The proposed dense correspondence algorithms will be quantitatively evaluated on these benchmarks. However, each of these benchmarks has its own limitations. We propose to develop a new benchmark for general correspondence estimation. It can be used for performance evaluation and automatic parameter selection for dense correspondence algorithms developed by other researchers.Some of the preliminary findings in this proposal were published in proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2009, 2010, 2012 and 2013, the European Conference on Computer Vision (ECCV) 2012 and the IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2009 and 2013.