Research map / Architectures and efficiency / RNNs and normalization
Removing Cost Volumes from Optical Flow Estimators
Simon Kiefhaber, Stefan Roth, Simone Schaub-Meyer
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 |
|---|---|---|---|
| remove the cost volume from the start of training (never train with it) (ablation) | ResNet-based RAFT-style estimator · Spring | EPE: 0.68 -> 0.86 | Hurt |
| train with cost volume first, then cut it off (cut-off strategy) (ablation) | ResNet-based estimator · Spring | EPE: 0.86 -> 0.68 | Improved |
| fade out cost volume with dropout instead of hard cut-off (ablation) | ReCoVEr training · Spring | EPE: 0.68 vs 0.68 | No clear effect |
| delay cut-off point from TartanAir to FlyingThings stage (ablation) | ReCoVEr-M training · Spring | EPE: 0.86 -> 0.68 | Improved |
| delay cut-off further to the TSKH stage (ablation) | ReCoVEr-M training · Spring | EPE: 0.68 -> 0.71 | Hurt |
| replace ResNet-34 context network with ConvNeXt-t (ablation) | ReCoVEr · Spring | EPE: 0.68 -> 0.51 | Improved |
| replace ResNet-34 context network with MobileNetV3-L (ablation) | ReCoVEr · Spring | EPE: 0.68 -> 0.99 | Hurt |
| remove cost volume (ReCoVEr-L vs SEA-RAFT) at inference | ReCoVEr-L · Spring total EPE | EPE: 0.54 -> 0.51 | Improved |
| remove cost volume at inference | ReCoVEr-L · Spring unmatched region | EPE: 7.25 -> 6.51 | Mixed |
| remove cost volume at inference | ReCoVEr-L · Out-of-domain Middlebury, Monkaa | EPE: Monkaa Final 2.41 vs 2.46 | No clear effect |
| downsample 2x then upsample flow | most optical flow methods · Spring vs Sintel/Monkaa | EPE: SEA-RAFT Spring 0.54 -> 0.41; Sintel Clean (0.43) -> (1.22) | Mixed |
| increase refinement iterations beyond 4 (non-ResNet models) (ablation) | ReCoVEr-S and ReCoVEr-L · Spring subset | EPE | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RAFT — RAFT: Recurrent All-Pairs Field Transforms for Optical Flow (ECCV 2020) · better on 1 of 1 shared result
- PWC-Net — PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume (CVPR 2018) · 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
- MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation (ICCV 2025 · similarity 0.90)
- Efficient All-Pairs Correlation Volume Sampling for Optical Flow Estimation (CVPR 2026 · similarity 0.89)
- MemFlow: Optical Flow Estimation and Prediction with Memory (CVPR 2024 · similarity 0.87)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 2024 · similarity 0.86)
- FlowDiffuser: Advancing Optical Flow Estimation with Diffusion Models (CVPR 2024 · similarity 0.85)
- DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework (CVPR 2025 · 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.
