Video action recognition with R(2+1)D
An R(2+1)D-18 network (pretrained on Kinetics-400, fine-tuned on UCF101) names the action in a video from 101 classes. The page samples five 16-frame clips across the video, runs them through the network in your browser and averages the predictions.
93.81% top-1 / 99.34% top-5
UCF101 split 1, 3,783 test videos
+39.4 points over training from scratch
Input
Works best on a few seconds of one clear action, such as a sport or playing an instrument. The 101 classes come from UCF101.
Output
The 16 frames of the middle clip as the model sees them (112×112), every 2nd frame shown.