Scene flow estimation based on Light Field
Le Wan, Xudong Zhang · 2022 7th International Conference on Intelligent Computing and Signal Processing (ICSP) · 2022
Light field camera can sample multiple views of the same scene with a single exposure, which has unique advantages in dynamic scene analysis. Existing scene flow estimation algorithms ignore the processing of image noise in natural scenes. To solve this problem, a scene estimation algorithm is designed using the light field focal stack. The bidirectional defocusing in the light field focal stack selects the optimal depth estimation cost and integrates noise processing into the depth estimation. The depth map is applied to the optical flow estimation through full focusing, so that the light rays in all directions of the spatial point are gathered into their respective angular domains, and the complete structural information and direction information of the spatial point can be obtained. When the spatial pixel points are matched, the constraints of the light in different directions of the spatial points are added, which greatly improves the matching accuracy of the algorithm. The anti-noise cost of depth estimation is diffused into the scene flow through all-focus, so that the scene flow algorithm has strong anti-noise performance. The experimental results show that the proposed method has lower sensitivity to noise.