Action Recognition on Kinetics-400 (RGB/Flow Stream Accuracy)
93.85Top-5 AccuracyMViT-B (16 × 4)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| MViT-B (16 × 4)Input=R, GFLOPs × views=70.80 × 5, Params (M)=36.62026.03 | 93.85 | 78.85 | — | — | |
| X3D-LInput=R, GFLOPs × views=26.64 × 30, Params (M)=6.22026.03 | 93.31 | 77.44 | — | — | |
| Two-Stream S3D-GInput=R+OF, GFLOPs × views=143.0 × N/A, Params (M)=23.12026.03 | 93 | — | — | 77.2 | |
| CAKE (ODConv3D)Input=R, GFLOPs × views=2.96 × 30, Params (M)=6.42026.03 | 92.3 | 73.33 | 44.71 | 74.2 | |
| Two-Stream I3DInput=R+OF, GFLOPs × views=216.0 × N/A, Params (M)=25.02026.03 | 92 | — | — | 75.7 | |
| SlowFast R50Input=R+OF, GFLOPs × views=36.69 × 30, Params (M)=34.52026.03 | 91.89 | — | — | 75.34 | |
| X3D-S (Baseline)Input=R, GFLOPs × views=1.96 × 30, Params (M)=3.82026.03 | 91.27 | 73.33 | — | — | |
| CAKE (Conv3D)Input=R, GFLOPs × views=2.85 × 30, Params (M)=6.12026.03 | 91.27 | 73.33 | 43.21 | — | |
| I3DInput=R, GFLOPs × views=108.0 × N/A, Params (M)=12.02026.03 | 90.3 | 71.1 | — | — | |
| Slow R50 (4 × 16)Input=R, GFLOPs × views=27.55 × 30, Params (M)=32.42026.03 | 90.18 | 72.4 | — | — | |
| C2D R50Input=R, GFLOPs × views=25.89 × 30, Params (M)=24.32026.03 | 89.68 | 71.46 | — | — | |
| Teacher X3D-SInput=OF, GFLOPs × views=1.96 × 30, Params (M)=3.82026.03 | — | — | 50.21 | — | |
| TSM R50Input=R, GFLOPs × views=65.0 × 10, Params (M)=24.32026.03 | — | 74.7 | — | — |