Reproduce “U$^{2}$Flow: Uncertainty-Aware Unsupervised Optical Flow Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xunpei Sun, Wenwei Lin, Yi Chang, Gang Chen
U²Flow introduces a recurrent unsupervised framework that jointly estimates dense optical flow and per-pixel uncertainty by deriving uncertainty supervision from augmentation consistency. The method integrates this learned uncertainty into the network for adaptive flow refinement, achieving a 6.13% Fl-all on KITTI-2015 and demonstrating superior uncertainty reliability.
Key insights
- Uncertainty supervision can be effectively generated in an unsupervised setting by measuring predictive inconsistencies under diverse data augmentations, enabling reliable uncertainty estimation without ground truth labels.
- Actively integrating learned uncertainty into the network architecture, through mechanisms like feature modulation and adaptive loss weighting, significantly enhances both optical flow accuracy and overall system robustness.
- A decoupled training strategy, separating the learning objectives for optical flow and uncertainty, is crucial for achieving stable and high-performing unsupervised joint estimation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.