Optical Flow Estimation on Sintel Clean (test)
0.79EPEGeoViT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GeoViTTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=true2025.12 | 0.79 | — | |
| MegaFlowTemporal Context=Multi-frame2026.03 | 0.91 | 4.84 | |
| WAFT-DINOv3-a2Temporal Context=Multi-frame2026.03 | 0.95 | 5.52 | |
| GG-SSMMulti-frame inputs=true2024.12 | 0.97 | — | |
| VideoFlow-MOFMulti-frame inputs=true2024.12 | 0.99 | — | |
| VideoFlow-MOFTemporal Context=Multi-frame2026.03 | 0.99 | 5.48 | |
| SAM-FlowTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1 | — | |
| SAMFlowTemporal Context=Multi-frame2026.03 | 1 | 5.25 | |
| VideoFlow-BOFMulti-frame inputs=true2024.12 | 1.01 | — | |
| VideoFlow-BOFTemporal Context=Multi-frame2026.03 | 1.01 | 5.61 | |
| FlowDiffuserTemporal Context=Multi-frame2026.03 | 1.02 | 5.57 | |
| GMFlow+Training Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.03 | — | |
| DPFlowTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.04 | — | |
| StreamFlowTemporal Context=Multi-frame2026.03 | 1.04 | 6 | |
| MemFlowMulti-frame inputs=true2024.12 | 1.05 | — | |
| TransFlowMulti-frame inputs=true2024.12 | 1.06 | — | |
| FlowFormer++2024.12 | 1.07 | — | |
| FlowFormer++Training Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.07 | — | |
| MemFlow-TTemporal Context=Multi-frame2026.03 | 1.08 | 6.02 | |
| CrocoFlowTemporal Context=Two-frame2026.03 | 1.09 | 6.3 | |
| MatchFlowWarm-start strategy=RAFT multi-frame2024.12 | 1.16 | — | |
| FlowFormerTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.16 | — | |
| FlowFormerTemporal Context=Two-frame2026.03 | 1.16 | 6.44 | |
| AnyFlowTemporal Context=Multi-frame2026.03 | 1.21 | 7.32 | |
| SKFlowWarm-start strategy=RAFT multi-frame2024.12 | 1.28 | — | |
| SKFlowTraining Data=C + T + S + K + H, Warm-start=true, Tiling technique=false2025.12 | 1.28 | — | |
| SEA-RAFT (L)Temporal Context=Two-frame2026.03 | 1.31 | 7.9 | |
| RPKNetTemporal Context=Two-frame2026.03 | 1.32 | 7.28 | |
| GMAWarm-start strategy=RAFT multi-frame2024.12 | 1.39 | — | |
| GMFlowNet2024.12 | 1.39 | — | |
| GMATraining Data=C + T + S + K + H, Warm-start=true, Tiling technique=false2025.12 | 1.39 | — | |
| GMFlowNetTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.39 | — | |
| GMATemporal Context=Two-frame2026.03 | 1.39 | 7.6 | |
| GMATraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.4 | — | |
| CRAFTTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.45 | — | |
| S-FlowTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.5 | — | |
| RAFT (warm-start)Training Data=C + T + S + K + H2020.03 | 1.61 | — | |
| RAFTWarm-start strategy=RAFT multi-frame2024.12 | 1.61 | — | |
| RAFTTraining Data=C + T + S + K + H, Warm-start=true, Tiling technique=false2025.12 | 1.61 | — | |
| RAFTTemporal Context=Two-frame2026.03 | 1.61 | 9.28 | |
| FM-RAFTTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.72 | — | |
| GMFlow2024.12 | 1.74 | — | |
| GMFlowTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.74 | — | |
| DDVMTemporal Context=Two-frame2026.03 | 1.75 | 12.22 | |
| RAFT (2-view)Training Data=C + T + S + K + H2020.03 | 1.94 | — | |
| RAFTTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 1.94 | — | |
| RAFT (2-view)Training Data=C + T + S/K2020.03 | 2.08 | — | |
| RAFT-ftLearning paradigm=Supervised, Fine-tuning=true2020.12 | 2.42 | — | |
| MaskFlowNetTraining Data=C + T + S + K + H2020.03 | 2.52 | — | |
| MaskFlowNetTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 2.52 | — | |
| MR-Flow2017.05 | 2.53 | — | |
| MRFlow2018.02 | 2.53 | — | |
| VCN-ftSupervision Type=Supervised2020.06 | 2.81 | — | |
| VCNTraining Data=C + T + S + K + H2020.03 | 2.81 | — | |
| VCNTraining Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 2.81 | — | |
| FlowFields++2020.03 | 2.94 | — | |
| FF+2017.05 | 3.1 | — | |
| PWC-FusionMulti-frame inputs=true2024.12 | 3.43 | — | |
| PWC-Net+Training Data=C + T + S + K + H2020.03 | 3.45 | — | |
| PWC-Net+Training Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 3.45 | — | |
| LiteFlowNet2Training Data=C + T + S + K + H2020.03 | 3.48 | — | |
| LiteFlowNet2Training Data=C + T + S + K + H, Warm-start=false, Tiling technique=false2025.12 | 3.48 | — | |
| DCFlow2018.02 | 3.54 | — | |
| DF2017.05 | 3.57 | — | |
| ScopeFlowTraining Data=C + T + S/K2020.03 | 3.59 | — | |
| SelFlow-ftSupervision Type=Supervised, Training Domain Condition (Labeled/Unlabeled/Related evaluation data)=Related evaluation data2020.06 | 3.74 | — | |
| FlowFields2020.03 | 3.75 | — | |
| FlowFields2018.02 | 3.78 | — | |
| IRR-PWCTraining Data=C + T + S/K2020.03 | 3.84 | — | |
| IRR-PWC+ftLearning paradigm=Supervised, Fine-tuning=true2020.12 | 3.84 | — | |
| PWC-Net-ftSupervision Type=Supervised2020.06 | 3.86 | — | |
| PWC-Net+ftLearning paradigm=Supervised, Fine-tuning=true2020.12 | 3.86 | — | |
| PWC-NetTemporal Context=Two-frame2026.03 | 3.86 | 25.29 | |
| FlowNet2fine-tuning=false2018.02 | 3.96 | — | |
| FlowNet2Supervision Type=Supervised2020.06 | 3.96 | — | |
| FlowNet2Training Data=C + T2020.03 | 3.96 | — | |
| EpicFlow2018.02 | 4.12 | — | |
| FlowNet2fine-tuning=true2018.02 | 4.16 | — | |
| FlowNet2-ftSupervision Type=Supervised2020.06 | 4.16 | — | |
| FlowNet2Training Data=C + T + S/K2020.03 | 4.16 | — | |
| Devonfine-tuning=true2018.02 | 4.34 | — | |
| PWC-Netfine-tuning=true2018.02 | 4.39 | — | |
| LiteFlowNetfine-tuning=true2018.02 | 4.54 | — | |
| UPFlowLearning paradigm=Unsupervised2020.12 | 4.68 | — | |
| ARFlowLearning paradigm=Unsupervised2020.12 | 4.78 | — | |
| HD3Training Data=C + T + S/K2020.03 | 4.79 | — | |
| LiteFlowNet+ftLearning paradigm=Supervised, Fine-tuning=true2020.12 | 4.86 | — | |
| UFlow-trainSupervision Type=Unsupervised2020.06 | 5.21 | — | |
| UFlowLearning paradigm=Unsupervised2020.12 | 5.21 | — | |
| SimFlowLearning paradigm=Unsupervised2020.12 | 5.92 | — | |
| STFlowLearning paradigm=Unsupervised2020.12 | 6.12 | — | |
| DDFlowSupervision Type=Unsupervised2020.06 | 6.18 | — | |
| DDFlowLearning paradigm=Unsupervised2020.12 | 6.18 | — | |
| SelFlowSupervision Type=Unsupervised, Additional Modalities (MDM, SDM, MF)=Multi-frame flow (MF), Training Domain Condition (Labeled/Unlabeled/Related evaluation data)=Related evaluation data2020.06 | 6.56 | — | |
| SelFlowLearning paradigm=Unsupervised2020.12 | 6.56 | — | |
| SPyNetFine-tuned=true, Time (s)=0.0692016.11 | 6.64 | — | |
| SpyNetfine-tuning=true2018.02 | 6.64 | — | |
| SpyNet+ftLearning paradigm=Supervised, Fine-tuning=true2020.12 | 6.64 | — | |
| SPyNetFine-tuned=false, Time (s)=0.0692016.11 | 6.69 | — | |
| SpyNetfine-tuning=false2018.02 | 6.69 | — |