Research map / Architectures and efficiency / RNNs and normalization
Disentangling Architecture and Training for Optical Flow
Deqing Sun, Charles Herrmann, Fitsum Reda, Michael Rubinstein, David J. Fleet, William T. Freeman
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 |
|---|---|---|---|
| pre-train on AutoFlow instead of FlyingChairs (ablation) | PWC-Net · Sintel final (pre-training) | AEPE: 4.79 -> 3.05 (with GC and OneCycle, 1.2M) | Improved |
| add gradient clipping (ablation) | IRR-PWC · KITTI (pre-training) | Fl-all: 26.95% -> 18.31% | Improved |
| add gradient clipping (ablation) | RAFT · Sintel final (pre-training) | AEPE: 3.36 -> 3.20 | Improved |
| use OneCycle instead of piecewise learning rate (ablation) | RAFT · KITTI (pre-training) | Fl-all: 17.95% -> 15.32% | Improved |
| train longer (1.2M to 6.2M iterations) (ablation) | PWC-Net · KITTI (pre-training) | Fl-all: 18.74% -> 16.29% | Improved |
| train RAFT longer (0.8M to 3.2M iterations) (ablation) | RAFT · Sintel final (pre-training) | AEPE: 2.57 -> 2.41 | Improved |
| raise gradient clipping threshold from 1.0 to 2.0 (ablation) | PWC-Net · KITTI (pre-training) | Fl-all: 18.74% -> 21.25% | Hurt |
| add weight decay, vertical flip or random erasing (ablation) | PWC-Net · Sintel (pre-training) | AEPE: 2.43 -> 2.31 (weight decay 1e-8), 2.46 (1e-7) | No clear effect |
| add gradient clipping in fine-tuning (ablation) | PWC-Net · Sintel val final (Lv split, fine-tuned) | AEPE: 4.98 -> 3.52 | Improved |
| use OneCycle instead of piecewise LR in fine-tuning (ablation) | PWC-Net · Sintel val final (Lv split, fine-tuned) | AEPE: 3.77 -> 3.52 | Improved |
| add AutoFlow to fine-tuning mixture (ablation) | PWC-Net · Sintel val final (Lv split, fine-tuned) | AEPE: 3.52 -> 3.25 | Improved |
| add AutoFlow to fine-tuning mixture (ablation) | PWC-Net · KITTI 2015 val (fine-tuned) | Fl-all: 4.76% -> 4.89% | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- SeparableFlow — Separable Flow: Learning Motion Cost Volumes for Optical Flow Estimation (ICCV 2021) · better on 1 of 3 shared results
- RAFT — RAFT: Recurrent All-Pairs Field Transforms for Optical Flow (ECCV 2020) · better on 3 of 3 shared results
- PWC-Net — PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume (CVPR 2018) · better on 3 of 3 shared results
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
- Models Matter, So Does Training: An Empirical Study of CNNs for Optical Flow Estimation (TPAMI 2019 · similarity 0.90)
- MS-RAFT+: High Resolution Multi-Scale RAFT (IJCV 2023 · similarity 0.89)
- PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume (CVPR 2018 · similarity 0.89)
- RAFT: Recurrent All-Pairs Field Transforms for Optical Flow (ECCV 2020 · similarity 0.88)
- "SEA-RAFT: Simple, Efficient, Accurate RAFT for Optical Flow" (ECCV 2024 · similarity 0.88)
- Context-Aware Iteration Policy Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.85)
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.
