Action Recognition on Kinetics-400 Initial (old) (test)
28.2Clip Error RateTimeSformer (Original)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| TimeSformer (Original)speed (clip per second)=<24.0, Backbone=TimeSformer2021.06 | 28.2 | — | 23.8 | |
| Temporal Differencespeed (clip per second)=≥81.0, Backbone=TimeSformer2021.06 | 32.3 | — | 26.9 | |
| Random Selectionspeed (clip per second)=≥81.0, Backbone=TimeSformer2021.06 | 34.3 | — | 28.2 | |
| IA-RED²speed (clip per second)=≥81.0, Backbone=TimeSformer2021.06 | 35.3 | — | 29.1 | |
| Attention Mapspeed (clip per second)=≥81.0, Backbone=TimeSformer2021.06 | 38.1 | — | 31.4 | |
| Non-local NNArchitecture=Non-local NN2018.11 | — | 77.7 | — | |
| Two-stream (2+1)DStream=Two-stream, Architecture=(2+1)D2018.11 | — | 75.4 | — | |
| Two-stream I3DStream=Two-stream, Architecture=I3D2018.11 | — | 74.1 | — | |
| Two-stream S3D-GStream=Two-stream, Architecture=S3D-G2018.11 | — | 77.2 | — |