A Closer Look at Spatiotemporal Convolutions for Action Recognition

Benchmark Model Rank Results
action-recognition-in-videos-on-ucf101R[2+1]D-TwoStream (Kinetics pretrained)#123-fold Accuracy: 97.3
action-recognition-in-videos-on-ucf101R[2+1]D-RGB (Kinetics pretrained)#173-fold Accuracy: 96.8
action-recognition-in-videos-on-ucf101R[2+1]D-Flow (Kinetics pretrained)#243-fold Accuracy: 95.5
action-recognition-in-videos-on-ucf101R[2+1]D-TwoStream (Sports-1M pretrained)#283-fold Accuracy: 95
action-recognition-in-videos-on-ucf101R[2+1]D-RGB (Sports-1M pretrained)#363-fold Accuracy: 93.6
action-recognition-in-videos-on-ucf101R[2+1]D-Flow (Sports-1M pretrained)#383-fold Accuracy: 93.3
action-recognition-in-videos-on-hmdb-51R[2+1]D-TwoStream (Kinetics pretrained)#17Average accuracy of 3 splits: 78.7
action-recognition-in-videos-on-hmdb-51R[2+1]D-Flow (Kinetics pretrained)#24Average accuracy of 3 splits: 76.4
action-recognition-in-videos-on-hmdb-51R[2+1]D-RGB (Kinetics pretrained)#27Average accuracy of 3 splits: 74.5
action-recognition-in-videos-on-hmdb-51R[2+1D]D-TwoStream (Sports1M pretrained)#30Average accuracy of 3 splits: 72.7
action-recognition-in-videos-on-hmdb-51R[2+1]D-Flow (Sports1M pretrained)#35Average accuracy of 3 splits: 70.1
action-recognition-in-videos-on-hmdb-51R[2+1]D-RGB (Sports1M pretrained)#39Average accuracy of 3 splits: 66.6
action-classification-on-kinetics-400R[2+1]D-Flow (Sports-1M pretrain)#149Acc@1: 75.4Acc@5: 91.9
action-classification-on-kinetics-400R[2+1]D-RGB (Sports-1M pretrain)#154Acc@1: 74.3Acc@5: 91.4
action-classification-on-kinetics-400R[2+1]D-Two-Stream#157Acc@1: 73.9Acc@5: 90.9
action-classification-on-kinetics-400R[2+1]D#163Acc@1: 72Acc@5: 90
action-classification-on-kinetics-400R[2+1]D-RGB#164Acc@1: 72Acc@5: 90
action-classification-on-kinetics-400R[2+1]D-Flow#172Acc@1: 67.5Acc@5: 87.2