Research map / Architectures and efficiency / CNN design
CRAFT: Cross-Attentional Flow Transformer for Robust Optical Flow
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use CRAFT's Semantic Smoothing Transformer and Cross-Frame Attention (ablation) | RAFT-style optical flow on KITTI-15 · KITTI-15 test | Fl-fg (%): CRAFT: 5.85; without SS trans: 6.41; without CFA: 6.15 | Helped |
| add the GMA module (ablation) | CRAFT on KITTI-15 · KITTI-15 test | Fl-fg (%): 5.85 -> 6.21 | Helped |
| use CRAFT instead of GMA | optical flow estimation · Things-Clean validation, GT motion >30 | AEPE: 7.71 -> 6.64 | Helped |
| use CRAFT instead of GMA | optical flow estimation · Things-Final validation, GT motion >30 | AEPE: 6.76 -> 6.11 | Helped |
| apply Semantic Smoothing Transformer to both frames instead of Frame 2 only (ablation) | CRAFT · optical flow matching | flow accuracy | Hurt |
| add relative position bias to the Semantic Smoothing Transformer (ablation) | CRAFT · optical flow estimation | flow accuracy | Helped |
| normalize the whole correlation volume (ablation) | CRAFT · optical flow estimation | flow accuracy | Helped |
| reduce GMA modes from 4 to 2 (ablation) | CRAFT · Sintel leaderboard Final, s40+ | AEPE: 13.27 -> 13.66 | Hurt |
| add mild ShiftAug during training (ablation) | CRAFT-m2 · Sintel leaderboard Final | AEPE: 2.42 -> 2.51 | Hurt |
| add mild ShiftAug during training (ablation) | CRAFT-m2 · Sintel leaderboard Clean | AEPE: 1.44 -> 1.40 | Helped |
| add aggressive ShiftAug during fine-tuning (ablation) | CRAFT · Chairs under image shifting attack, Δx=400 | AEPE: 336.9 -> 32.0 | Helped |
| add aggressive ShiftAug during fine-tuning (ablation) | GMA · Chairs under image shifting attack, Δx=400 | AEPE: 478.1 -> 124.3 | Helped |
Closest papers by meaning
- TransFlow: Transformer As Flow Learner (CVPR 2023 · similarity 0.89)
- Optical Flow Estimation from a Single Motion-blurred Image (AAAI 2021 · similarity 0.88)
- Learning Optical Flow From a Few Matches (CVPR 2021 · similarity 0.88)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.86)
- FlowFeat: Pixel-Dense Embedding of Motion Profiles (NeurIPS 2025 · similarity 0.86)
- FlowDiffuser: Advancing Optical Flow Estimation with Diffusion Models (CVPR 2024 · similarity 0.86)
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.