Unsupervised Learning of Optical Flow With CNN-based Non-Local Filtering.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Estimating optical flow is crucial for computer vision and image processing.
• Unsupervised deep learning methods for optical flow often suffer from over-smoothing and occlusion issues.
• Current unsupervised approaches lag behind supervised methods in performance.

Purpose of the Study:

• To propose a novel unsupervised method for accurate optical flow estimation.
• To address limitations of existing methods, specifically over-smoothing and occlusion handling.
• To achieve state-of-the-art performance in unsupervised optical flow estimation.

Main Methods:

• A novel post-processing term using a CNN-based non-local approach refines optical flow, reducing noise and blur at motion boundaries.
• An effective loss function incorporates a symmetrical energy formulation to detect and mitigate occlusion effects.
• The method is trained end-to-end for unsupervised optical flow estimation.

Main Results:

• The proposed method effectively refines optical flow by learning spatial dependencies over large neighborhoods.
• Occlusion maps generated from bi-directional flows are integrated into the loss function, reducing occlusion influence.
• State-of-the-art results were achieved on FlyingChairs, MPI-Sintel, and KITTI datasets.

Conclusions:

• The novel unsupervised method significantly improves optical flow estimation accuracy.
• The combination of non-local refinement and occlusion-aware loss effectively tackles key challenges.
• The approach demonstrates superior performance compared to existing unsupervised methods.