Non-Learning Low-Light Stereo Vision
Researchers have developed a new stereo vision framework designed to work effectively in low-light conditions, even with severely noisy images. This non-learning approach utilizes the Field of Junctions (FoJ) to identify stable visual features for cost volume construction, while ignoring fine textures that are indistinguishable from noise. The system then employs a boundary-aware Semi-Global Matching (SGM) algorithm that adapts its smoothness penalties to preserve accurate disparity discontinuities, resulting in a sparse disparity map that outperforms recent stereo algorithms on benchmark datasets. AI
IMPACT This novel stereo vision method could improve performance in low-light and noisy environments for applications like robotics and autonomous driving.