Research map / Architectures and efficiency / RNNs and normalization
FlowFormer: A Transformer Architecture for Optical Flow
Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, Hongsheng Li
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 |
|---|---|---|---|
| replace RAFT multi-level cost retrieval decoder with latent cost tokens plus cost memory decoder (K=8, D=128) (ablation) | FlowFormer (RAFT-based ablation) · Sintel train clean | AEPE: 1.53 -> 1.44 | Improved |
| increase latent tokens from K=4,D=32 to K=8,D=128 (ablation) | FlowFormer (RAFT-based ablation) · Sintel train clean | AEPE: 1.66 -> 1.44 | Improved |
| replace CNN image encoder with Twins trained from scratch (ablation) | FlowFormer (RAFT-based ablation) · Sintel train final | AEPE: 2.80 -> 2.86 | Hurt |
| use ImageNet-pretrained Twins instead of CNN image encoder (ablation) | FlowFormer (RAFT-based ablation) · Sintel train clean | AEPE: 1.44 -> 1.29 | Improved |
| use ImageNet-pretrained Twins instead of CNN image encoder (ablation) | FlowFormer (RAFT-based ablation) · KITTI-15 train | F1-all (%): 17.64 -> 16.16 | Improved |
| add intra-cost-map attention only (ablation) | FlowFormer (RAFT-based ablation) · Sintel train final | AEPE: 2.72 -> 2.89 | Hurt |
| add one AGT layer (intra + inter attention) (ablation) | FlowFormer (RAFT-based ablation) · Sintel train clean | AEPE: 1.29 -> 1.20 | Improved |
| stack three AGT layers instead of none (ablation) | FlowFormer (RAFT-based ablation) · Sintel train final | AEPE: 2.72 -> 2.57 | Improved |
| stack three AGT layers instead of none (ablation) | FlowFormer (RAFT-based ablation) · KITTI-15 train | F1-all (%): 16.16 -> 15.15 | Improved |
| use FlowFormer (small) with GMA CNN encoder versus GMA (ablation) | FlowFormer (RAFT-based ablation) · Sintel train clean | AEPE: 1.30 -> 1.20 | Improved |
| enlarge GMA to GMA-L (17.0M parameters) to match size | FlowFormer (RAFT-based ablation) · Sintel train clean | AEPE: 1.30 -> 1.33 | No clear effect |
| give GMA a pretrained Twins encoder | FlowFormer (RAFT-based ablation) · KITTI-15 train | F1-epe: 4.69 -> 4.98 | Hurt |
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.
- RAFT — RAFT: Recurrent All-Pairs Field Transforms for Optical Flow (ECCV 2020) · better on 7 of 7 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
- FlowFormer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation (CVPR 2023 · similarity 0.89)
- FlowerFormer: Empowering Neural Architecture Encoding using a Flow-aware Graph Transformer (CVPR 2024 · similarity 0.88)
- FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation (ECCV 2026 · similarity 0.86)
- FlowNAS: Neural Architecture Search for Optical Flow Estimation (IJCV 2023 · similarity 0.86)
- AutoFormer: Searching Transformers for Visual Recognition (ICCV 2021 · similarity 0.85)
- Flowformer: Linearizing Transformers with Conservation Flows (ICML 2022 · 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.
