Paper Overview
- Research area: Computer Vision (CV)
- Authors: Yixuan Luo, Feng Qiao, Zhexiao Xiong, Yanjing Li, Nathan Jacobs
- Published: 2026-03-23
- arXiv: 2603.22270
Motivation
Optical flow estimation is a fundamental problem in computer vision, yet reliance on expensive ground-truth annotations limits the scalability of supervised approaches. Unsupervised and semi-supervised methods alleviate this issue but often suffer from unreliable supervision signals based on brightness constancy and smoothness assumptions, leading to inaccurate motion estimation in complex real-world scenarios.
Proposed Approach
To overcome these limitations, the authors introduce GenOpticalFlow, a novel framework that synthesizes large-scale, perfectly aligned frame–flow data pairs for supervised optical flow training without human annotations.
Key components:
1. Pseudo optical flow generation — a pre-trained depth estimation network produces pseudo optical flows, which serve as conditioning inputs for a next-frame generation model. 2. Next-frame generation — the generative model produces high-fidelity, pixel-aligned subsequent frames, enabling large-scale synthetic data with precise motion correspondences. 3. Inconsistent pixel filtering — a strategy that identifies and removes unreliable pixels from generated frames, effectively improving fine-tuning performance on real-world datasets.
Summary
GenOpticalFlow demonstrates a scalable, annotation-free route to supervised-quality optical flow training data, combining generative frame synthesis with filtering to improve robustness on real-world benchmarks.
--- *Auto-collected on 2026-03-25*