Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
MegaFlow: Zero-Shot Large Displacement Optical Flow
Dingxi Zhang, Fangjinhua Wang, Marc Pollefeys, Haofei Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add VGGT pretrained prior to a 12-layer backbone (ablation) | MegaFlow · Zero-shot Sintel Final | EPE: 3.31 -> 2.23 | Improved |
| scale transformer from 6 to 12 layers without pre-training (ablation) | MegaFlow from scratch · Zero-shot Sintel Clean | EPE: 2.05 -> 2.13 | Hurt |
| keep transformer trainable instead of frozen (ablation) | MegaFlow · Zero-shot KITTI | Fl-all: 14.7 -> 10.9 | Improved |
| add CNN feature fusion (ablation) | MegaFlow · Zero-shot KITTI | Fl-epe: 4.71 -> 3.00 | Improved |
| add temporal attention with 4-frame input (ablation) | MegaFlow · Zero-shot Sintel Final | EPE: 1.99 -> 1.83 | Improved |
| increase input frames from 2 to 4 with temporal attention (ablation) | MegaFlow · Zero-shot Sintel Final | EPE: 2.08 -> 1.83 | Improved |
| increase input frames from 2 to 6 with temporal attention (ablation) | MegaFlow · Zero-shot KITTI | Fl-epe: 3.00 -> 3.67 | Hurt |
| add TartanAir training stage (ablation) | MegaFlow · Zero-shot Sintel Clean | EPE: 0.94 -> 0.85 | Improved |
| add 2-frame FlyingThings warmup stage (ablation) | MegaFlow · Zero-shot KITTI | Fl-all: 12.2 -> 11.1 | Improved |
| increase refinement iterations from 1 to 8 (ablation) | MegaFlow · Zero-shot KITTI | Fl-all: 19.03 -> 11.14 | Improved |
| increase refinement iterations beyond 8 (K=10) (ablation) | MegaFlow · Zero-shot Sintel Final | EPE: 1.83 -> 1.85 | No clear effect |
| fine-tune on Kubric after flow-mix training | MegaFlow · TAP-Vid mean | delta_avg: 73.6 -> 79.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CoTracker3 — CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos (ICCV 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MemFlow: Optical Flow Estimation and Prediction with Memory (CVPR 2024 · similarity 0.88)
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.88)
- SAMFlow: Eliminating Any Fragmentation in Optical Flow with Segment Anything Model (AAAI 2024 · similarity 0.87)
- Zero-Shot Monocular Scene Flow Estimation in the Wild (CVPR 2025 · similarity 0.87)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.87)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
