OIFlow: Occlusion-Inpainting Optical Flow Estimation by Unsupervised Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Occlusion presents a significant challenge in unsupervised optical flow learning.
• Current methods inadequately address occlusions, either by treating them uniformly or by exclusion.
• Occluded regions contain valuable information for accurate optical flow estimation.

Purpose of the Study:

• To develop an innovative occlusion-inpainting framework, termed OIFlow.
• To fully utilize occlusion regions for enhancing unsupervised optical flow learning.
• To improve the accuracy and robustness of optical flow estimation in the presence of occlusions.

Main Methods:

• Proposed a novel appearance-flow network for inpainting occluded optical flows using image content.
• Introduced a boundary dilated warp mechanism to manage occlusions arising from displacements beyond image boundaries.
• Evaluated the framework on benchmark datasets including Flying Chairs, KITTI, and MPI-Sintel.

Main Results:

• The OIFlow framework demonstrated significant performance improvements across multiple leading optical flow benchmark datasets.
• The occlusion-inpainting approach effectively integrated information from occluded regions.
• The boundary dilated warp successfully addressed occlusions caused by out-of-bounds displacements.

Conclusions:

• The proposed OIFlow framework offers a superior method for handling occlusions in unsupervised optical flow learning.
• Leveraging occlusion information through inpainting leads to substantial gains in optical flow estimation accuracy.
• This work advances the state-of-the-art in optical flow estimation by effectively addressing the critical occlusion problem.