actNet_mse:
similarity matrices based ActivityNet-trained model (with l2 loss)
actNet_gdl:
similarity matrices based ActivityNet-trained model (with gradient difference loss)
20bn_mse:
Similarity matrices based on 20BN-trained model (with l2 loss)
20bn_gdl:
Similarity matrices based on 20BN-trained model (with gradient difference loss)
20bn_gdl_optical_flow:
Similarity matrices based on 20BN-trained model (with gradient difference loss, 4th channel is dense optical flow of original image computed with OpenCV’s calcOpticalFlowFarneback())