Learning by Distillation: A Self-Supervised Learning Framework for Optical Flow Estimation

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Optical flow estimation is crucial for understanding motion in videos.
• Current supervised methods heavily rely on large labeled datasets, often pre-trained on synthetic data.
• Learning optical flow for occluded regions remains a significant challenge.

Purpose of the Study:

• To introduce DistillFlow, a novel knowledge distillation approach for learning optical flow.
• To enable effective self-supervised learning of optical flow from unlabeled data, including occluded pixels.
• To demonstrate the generalization capabilities of the proposed method.

Main Methods:

• DistillFlow employs multiple teacher models and a student model.
• Challenging transformations generate hallucinated occlusions and less confident predictions for the student.
• A self-supervised framework uses confident teacher predictions as annotations to guide the student model.

Main Results:

• DistillFlow achieves state-of-the-art unsupervised learning performance on KITTI and Sintel datasets.
• Self-supervised pre-training with DistillFlow provides excellent initialization for supervised fine-tuning.
• Fine-tuned models achieved top rankings on KITTI 2015 and outperformed existing methods on Sintel Final.

Conclusions:

• DistillFlow offers an effective alternative to traditional supervised learning paradigms for optical flow.
• The method demonstrates strong generalization across frameworks, correspondences, and datasets.
• The approach successfully learns optical flow for both non-occluded and occluded regions using unlabeled data.