.png){kind=icon}
.png){kind=icon}
Return
OMFlow: Optimizing optical flow via occlusion motion estimation
.png){kind=icon}
PRE
AI
W
W
D
L
X
J
DOI：10.1016/j.patrec.2026.03.018.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Occlusions pose a significant challenge to optical flow algorithms that heavily depend on local cues. The contemporary optical flow techniques focus on the feature enhancement of 4D correlation volume while ignoring the fact that occlusions' matches no longer exist in the volume. In this work, we propose an optical flow estimation model called OMFlow to recalculate the occlusion matching based on the motion information and improve the accuracy of correlation matching. Our work is the first to combine the pattern of motion with the correlation volume and replace the feature similarity comparison by estimating the potential matching of the occluded point. The obtained motion information helps to learn more fundamental correlations of objects and scenes with consistent displacements. Specifically, it introduces a multi-factor feature similarity judgment function to detect the occlusion resistant to noise interference and a byte track-based network to adjust the correlation volume of occluded points. OMFlow has good portability, and can easily serve existing methods as a plug-and-play component. It achieves 23.7% and 9.1% error reduction from GMA on Sintel clean pass and final pass. In the occlusion analysis, our method reduces the error of the occlusion area by 88.4% and 85.4%. The experiment shows that our method is state-of-the-art by modifying the corresponding relation of the occlusion.
Keywords:
Occlusion
Motion model
Correlation
Journal
IF:
3.3
Papers:
7.8K
Citations:
1.6W
Organization
Cited Papers