Action Recognition on Kinetics-400 (val)
87.5Top-1 AccAIM ViT-L/14
Evaluation results
| Method | Links | ||
|---|---|---|---|
| AIM ViT-L/14Strategy=PETL, Pretrain=CLIP, Views=32×1×3, GFLOPs=11208, Extra GFLOPs=3425, Param(M)=341, New Param(M)=382023.10 | 87.5 | 97.7 | |
| ST-Adapter ViT-L/14Strategy=PETL, Pretrain=CLIP, Views=32×1×3, GFLOPs=8248, Extra GFLOPs=322, Param(M)=319, New Param(M)=152023.10 | 87.2 | 97.6 | |
| ZeroI2V ViT-L/14Pretrain=CLIP, Views=32×1×3, GFLOPs=7783, Extra GFLOPs=0, Param(M)=304, New Param(M)=02023.10 | 87.2 | 97.6 | |
| X-CLIP ViT-L/14Strategy=full fine-tuning, Pretrain=CLIP, Views=8×3×4, GFLOPs=7896, Extra GFLOPs=107, Param(M)=420, New Param(M)=1162023.10 | 87.1 | 97.6 | |
| Text4Vis ViT-L/14Strategy=full fine-tuning, Pretrain=CLIP, Views=32×3×4, GFLOPs=19944, Param(M)=347, New Param(M)=432023.10 | 87.1 | 97.4 | |
| ZeroI2V ViT-L/14Pretrain=CLIP, Views=16×1×3, GFLOPs=3892, Extra GFLOPs=0, Param(M)=304, New Param(M)=02023.10 | 86.8 | 97.6 | |
| EVL ViT-L/14Strategy=PETL, Pretrain=CLIP, Views=8×1×3, GFLOPs=2022, Extra GFLOPs=76, Param(M)=362, New Param(M)=582023.10 | 86.3 | — | |
| ZeroI2V ViT-L/14Pretrain=CLIP, Views=8×1×3, GFLOPs=1946, Extra GFLOPs=0, Param(M)=304, New Param(M)=02023.10 | 86.3 | 97.4 | |
| MViTv2-L(↑312)Strategy=full fine-tuning, Pretrain=IN21K, Views=40×5×3, GFLOPs=42420, Param(M)=2182023.10 | 86.1 | 97 | |
| Swin-L (384↑)Pretrain=ImageNet-21K, Views=10 x 5, FLOPs=2107G, Param=200.0M, Resolution=384x3842021.06 | 84.9 | 96.7 | |
| ViViT-H/16x2Pretrain=JFT-300M, Views=4 x 3, FLOPs=8316G, Param=647.5M2021.06 | 84.8 | 95.8 | |
| Swin-L (384↑)Pretrain=ImageNet-21K, Views=4 x 3, FLOPs=2107G, Param=200.0M, Resolution=384x3842021.06 | 84.6 | 96.5 | |
| MTV-LStrategy=full fine-tuning, Pretrain=JFT, Views=32×3×4, GFLOPs=18050, Param(M)=8762023.10 | 84.3 | 96.3 | |
| ActionCLIP-B/16Strategy=full fine-tuning, Pretrain=CLIP, Views=32×3×10, GFLOPs=16893, Extra GFLOPs=13, Param(M)=142, New Param(M)=562023.10 | 83.8 | 97.1 | |
| ZeroI2V ViT-B/16Pretrain=CLIP, Views=32×1×3, GFLOPs=1688, Extra GFLOPs=0, Param(M)=86, New Param(M)=02023.10 | 83.7 | 96.4 | |
| ViViT-L/16x2 320Pretrain=JFT-300M, Views=4 x 3, FLOPs=3992G, Param=310.8M, Resolution=3202021.06 | 83.5 | 95.5 | |
| ViViT-L/16x2 FEStrategy=full fine-tuning, Pretrain=JFT, Views=32×3×1, GFLOPs=11940, Param(M)=3112023.10 | 83.5 | 94.3 | |
| ZeroI2V ViT-B/16Pretrain=CLIP, Views=16×1×3, GFLOPs=844, Extra GFLOPs=0, Param(M)=86, New Param(M)=02023.10 | 83.4 | 96.2 | |
| Swin-LPretrain=ImageNet-21K, Views=4 x 3, FLOPs=604G, Param=197.0M2021.06 | 83.1 | 95.9 | |
| VideoSwin-LStrategy=full fine-tuning, Pretrain=IN21K, Views=32×3×4, GFLOPs=7248, Param(M)=1972023.10 | 83.1 | 95.9 | |
| UniFormer-BStrategy=full fine-tuning, Pretrain=IN1K, Views=32×3×4, GFLOPs=3108, Param(M)=502023.10 | 83 | 95.4 | |
| ZeroI2V ViT-B/16Pretrain=CLIP, Views=8×1×3, GFLOPs=422, Extra GFLOPs=0, Param(M)=86, New Param(M)=02023.10 | 83 | 95.8 | |
| ViViT-L/16x2Pretrain=JFT-300M, Views=4 x 3, FLOPs=1446G, Param=310.8M2021.06 | 82.8 | 95.5 | |
| Swin-BPretrain=ImageNet-21K, Views=4 x 3, FLOPs=282G, Param=88.1M2021.06 | 82.7 | 95.5 | |
| ip-CSN-152Pretrain=IG-65M, Views=10 x 3, FLOPs=109G, Param=32.8M2021.06 | 82.5 | 95.3 | |
| TokenLearner 16at12 (L/16)Views=4x3, GFLOPs=45962021.11 | 82.1 | — | |
| ATAPretrain=IN-21K, Frames=32, Views=3 × 4, FLOPs (G)=792.9, Params (M)=121.82022.09 | 81.9 | 95.5 | |
| ATAPretrain=IN-21K, Frames=32, Views=3 × 1, FLOPs (G)=792.9, Params (M)=121.82022.09 | 81.4 | 95.5 | |
| ViViT-L/16x2 320Pretrain=ImageNet-21K, Views=4 x 3, FLOPs=3992G, Param=310.8M, Resolution=3202021.06 | 81.3 | 94.7 | |
| ViViT-LPretrain=IN-21K, Frames=16, Views=3 × 4, FLOPs (G)=3992, Params (M)=310.82022.09 | 81.3 | 94.7 | |
| LGD-3D Two-streamBackbone=ResNet-1012019.06 | 81.2 | 95.2 | |
| MViT-B, 64×3Pretrain=None, Views=3 x 3, FLOPs=455G, Param=36.6M2021.06 | 81.2 | 95.1 | |
| MViT-B, 64x3Views=3x3, GFLOPs=40952021.11 | 81.2 | 95.1 | |
| MViTv1-BFrames=64, Views=3 × 3, FLOPs (G)=455, Params (M)=36.62022.09 | 81.2 | 95.1 | |
| Mformer-HRPretrain=IN-21K, Frames=16, Views=3 × 10, FLOPs (G)=958.8, Params (M)=381.92022.09 | 81.1 | 95.2 | |
| ViT-B/16Strategy=full fine-tuning, Pretrain=CLIP, Views=8×1×3, GFLOPs=422, Extra GFLOPs=0, Param(M)=86, New Param(M)=02023.10 | 81 | 95.5 | |
| TimeSformer-LPretrain=ImageNet-21K, Views=1 x 3, FLOPs=2380G, Param=121.4M2021.06 | 80.7 | 94.7 | |
| TimeSformer-LViews=1x3, GFLOPs=71402021.11 | 80.7 | 94.7 | |
| TimeSformer-LPretrain=IN-21K, Frames=96, Views=3 × 1, FLOPs (G)=2380, Params (M)=121.42022.09 | 80.7 | 94.7 | |
| TimeSformer-LStrategy=full fine-tuning, Pretrain=IN21K, Views=96×3×1, GFLOPs=7140, Param(M)=1212023.10 | 80.7 | 94.7 | |
| ViViT-L/16x2Pretrain=ImageNet-21K, Views=4 x 3, FLOPs=1446G, Param=310.8M2021.06 | 80.6 | 94.7 | |
| Swin-SPretrain=ImageNet-1K, Views=4 x 3, FLOPs=166G, Param=49.8M2021.06 | 80.6 | 94.5 | |
| Swin-BPretrain=ImageNet-1K, Views=4 x 3, FLOPs=282G, Param=88.1M2021.06 | 80.6 | 94.6 | |
| ViViT-L/16x2Views=4x3, GFLOPs=173522021.11 | 80.6 | 94.7 | |
| TimeSformer-L+ATSViews=1x3, GFLOPs=35102021.11 | 80.5 | 94.6 | |
| X3D-XXLPretrain=None, Views=10 x 3, FLOPs=144G, Param=20.3M2021.06 | 80.4 | 94.6 | |
| X3D-XXLViews=10x3, GFLOPs=58232021.11 | 80.4 | 94.6 | |
| MViT-B, 32×3Pretrain=None, Views=1 x 5, FLOPs=170G, Param=36.6M2021.06 | 80.2 | 94.4 | |
| X-ViT (16x)Views=1x3, GFLOPs=4252021.11 | 80.2 | 94.7 | |
| MViTv1-BFrames=32, Views=1 × 5, FLOPs (G)=170, Params (M)=36.62022.09 | 80.2 | 94.4 | |
| Mformer-LPretrain=IN-21K, Frames=32, Views=3 × 10, FLOPs (G)=1185.1, Params (M)=109.12022.09 | 80.2 | 94.8 | |
| X-VIT+ATS (16x)Views=1x3, GFLOPs=2592021.11 | 80 | 94.6 | |
| SlowFast R101+NLPretrain=None, Views=10 x 3, FLOPs=234G, Param=59.9M2021.06 | 79.8 | 93.9 | |
| SlowFast 16x8 R101+NLViews=10x3, GFLOPs=70202021.11 | 79.8 | 93.9 | |
| ViT-B-VTNPretrain=IN-21K, Frames=16, Views=1 × 1, FLOPs (G)=4218, Params (M)=1142022.09 | 79.8 | 94.2 | |
| TimeSformer-HRPretrain=IN-21K, Frames=16, Views=3 × 1, FLOPs (G)=1703, Params (M)=121.42022.09 | 79.7 | 94.4 | |
| Mformer-BPretrain=IN-21K, Frames=16, Views=3 × 10, FLOPs (G)=369.5, Params (M)=109.12022.09 | 79.7 | 94.2 | |
| LGD-3D RGBBackbone=ResNet-1012019.06 | 79.4 | 94.4 | |
| HATNet2021.11 | 79.3 | — | |
| CorrNet-101Pretrain=None, Views=10 x 3, FLOPs=224G2021.06 | 79.2 | — | |
| I3D NLViews=10x3, GFLOPs=67002021.11 | 79.2 | — | |
| ip-CSN-152Views=10x3, GFLOPs=32702021.11 | 79.2 | 93.8 | |
| VidTr-LPretrain=IN-21K, Frames=32, Views=3 × 10, FLOPs (G)=3512022.09 | 79.1 | 93.9 | |
| Swin-TPretrain=ImageNet-1K, Views=4 x 3, FLOPs=88G, Param=28.2M2021.06 | 78.8 | 93.6 | |
| AIM ViT-B-14Strategy=PETL, Pretrain=IN21K, Views=8×1×3, GFLOPs=624, Extra GFLOPs=202, Param(M)=100, New Param(M)=142023.10 | 78.8 | — | |
| ViT-B-VTNPretrain=ImageNet-21K, Views=1 x 1, FLOPs=4218G, Param=11.04M2021.06 | 78.6 | 93.7 | |
| ZeroI2V ViT-B/16Pretrain=IN21K, Views=8×1×3, GFLOPs=422, Extra GFLOPs=0, Param(M)=86, New Param(M)=02023.10 | 78.6 | — | |
| MViTv1-BFrames=16, Views=1 × 5, FLOPs (G)=70.5, Params (M)=36.62022.09 | 78.4 | 93.5 | |
| TimeSformerPretrain=IN-21K, Frames=8, Views=3 × 1, FLOPs (G)=196, Params (M)=121.42022.09 | 78 | 93.7 | |
| ip-CSN-152Pretrain=None, Views=10 x 3, FLOPs=109G, Param=32.8M2021.06 | 77.8 | 92.8 | |
| NL I3DModality=RGB2018.12 | 77.7 | — | |
| SANModality=RGB+Flow+Audio2018.12 | 77.7 | — | |
| NL I3D RGBBackbone=ResNet-1012019.06 | 77.7 | 93.3 | |
| Three-stream SATTBackbone=Inception-ResNet-v2, Reference=Anet17 winner report2019.06 | 77.7 | 93.2 | |
| NL I3D-101Pretrain=ImageNet-1K, Views=10 x 3, FLOPs=359G, Param=61.8M2021.06 | 77.7 | 93.3 | |
| TSM R50Views=10x3, GFLOPs=108002021.11 | 77.7 | 93.3 | |
| SSAEnFrames=8+16, Backbone=ResNet-502021.05 | 77.5 | 93.3 | |
| S3D-GModality=RGB+Flow2018.12 | 77.2 | — | |
| S3D-G Two-streamBackbone=Inception2019.06 | 77.2 | 93 | |
| VideoPrompt ViT-B/16Strategy=PETL, Pretrain=CLIP, Views=16×5×12023.10 | 76.9 | 93.5 | |
| ST-Adapter ViT-B/16Strategy=PETL, Pretrain=IN21K, Views=8×1×3, GFLOPs=455, Extra GFLOPs=33, Param(M)=93, New Param(M)=72023.10 | 76.6 | — | |
| D3D+S3D-GModality=RGB+RGB2018.12 | 76.5 | — | |
| NL I3DFrames=128, Backbone=ResNet-502021.05 | 76.5 | 92.6 | |
| SSAFrames=16, Backbone=ResNet-502021.05 | 76.4 | 92.7 | |
| S3D-GModality=RGB+Flow2018.12 | 76.2 | — | |
| STMViews=10x3, GFLOPs=21002021.11 | 76.1 | 92.5 | |
| GCNetBackbone=ResNet-50, #params(M)=42.45, FLOPs(G)=39.352020.12 | 76 | 92.34 | |
| NLNetBackbone=ResNet-50, #params(M)=39.81, FLOPs(G)=59.602020.12 | 75.95 | 92.29 | |
| D3DModality=RGB2018.12 | 75.9 | — | |
| SSAFrames=8, Backbone=ResNet-502021.05 | 75.8 | 92.4 | |
| I3D Two-streamBackbone=Inception2019.06 | 75.7 | 92 | |
| R(2+1)DModality=RGB+Flow2018.12 | 75.4 | — | |
| R(2+1)D Two-streamBackbone=custom2019.06 | 75.4 | 91.9 | |
| EVL ViT-B/16Strategy=PETL, Pretrain=IN21K, Views=8×1×3, GFLOPs=454, Extra GFLOPs=32, Param(M)=115, New Param(M)=292023.10 | 75.4 | — | |
| GloREBackbone=ResNet-50, FLOPs(G)=28.90, Version=Lite version of C2D2020.12 | 75.12 | — | |
| Slow-onlyBackbone=ResNet-50, #params(M)=32.45, FLOPs(G)=39.292020.12 | 74.94 | 91.9 | |
| Three-stream late fusionBackbone=Inception-ResNet-v2, Reference=Anet17 winner report2019.06 | 74.9 | 91.6 | |
| S3D-GModality=RGB2018.12 | 74.7 | — | |
| S3D-G RGBBackbone=Inception2019.06 | 74.7 | 93.4 | |
| TEAViews=10x3, GFLOPs=6502021.11 | 74.7 | — |