This is a general-purpose action recognition composite model, consisting of encoder and decoder parts, trained on Kinetics-400 dataset. The encoder model uses Video Transformer approach with ResNet34 encoder. Please refer to the kinetics dataset specification to see list of action that are recognised by this composite model.
The action-recognition-0001-encoder model accepts video frame and produces embedding. Video frames should be sampled to cover ~1 second fragment (i.e. skip every second frame in 30 fps video).
name: "0", shape: [1x3x224x224] - An input image in the format [BxCxHxW], where:
Expected color order is BGR.
The model outputs a tensor with the shape [1x512x1x1], representing embedding of processed frame.
The action-recognition-0001-decoder model accepts stack of frame embeddings, computed by action-recognition-0001-encoder model.
The model outputs a tensor with the shape [bx400], each row is a logits vector of performed actions.
[*] Other names and brands may be claimed as the property of others.