Multiscale Vision Transformers

Benchmark Model Rank Results
action-classification-on-charadesMViT-B-24, 32x3 (Kinetics-600 pretraining)#10MAP: 47.7
action-classification-on-charadesMViT-B, 32x3 (Kinetics-600 pretraining)#12MAP: 47.1
action-classification-on-charadesMViT-B-24, 32x3 (Kinetics-400 pretraining)#13MAP: 46.3
action-classification-on-charadesMViT-B, 32x3 (Kinetics-400 pretraining)#16MAP: 44.3
action-classification-on-charadesMViT-B, 16x4 (Kinetics-600 pretraining)#18MAP: 43.9
action-classification-on-charadesMViT-B, 16x4 (Kinetics-400 pretraining)#24MAP: 40
action-recognition-on-ava-v2-2MViT-B-24, 32x3 (Kinetics-600 pretraining)#25mAP: 28.7
action-recognition-on-ava-v2-2MViT-B, 32x3 (Kinetics-500 pretraining)#27mAP: 27.5
action-recognition-on-ava-v2-2MViT-B, 64x3 (Kinetics-400 pretraining)#28mAP: 27.3
action-recognition-on-ava-v2-2MViT-B, 32x3 (Kinetics-400 pretraining)#30mAP: 26.8
action-recognition-on-ava-v2-2MViT-B, 16x4 (Kinetics-600 pretraining)#33mAP: 26.1
action-recognition-on-ava-v2-2MViT-B, 16x4 (Kinetics-400 pretraining)#34mAP: 24.5
action-classification-on-kinetics-600MViT-B-24, 32x3#30Top-1 Accuracy: 83.8Top-5 Accuracy: 96.3
action-classification-on-kinetics-600MViT-B, 32x3#33Top-1 Accuracy: 83.4Top-5 Accuracy: 96.3
action-classification-on-kinetics-600MViT-B, 16x4#38Top-1 Accuracy: 82.1Top-5 Accuracy: 95.7
action-recognition-in-videos-on-something-something-v2MViT-B-24, 32x3#46Top-1 Accuracy: 68.7Top-5 Accuracy: 91.5Parameters: 53.2M…
action-recognition-in-videos-on-something-something-v2MViT-B, 32x3(Kinetics600 pretrain)#52Top-1 Accuracy: 67.8Top-5 Accuracy: 91.3Parameters: 36.6…
action-recognition-in-videos-on-something-something-v2MViT-B, 16x4#73Top-1 Accuracy: 66.2Top-5 Accuracy: 90.2
action-classification-on-kinetics-400MViT-B, 64x3#79Acc@1: 81.2Acc@5: 95.1
action-classification-on-kinetics-400MViT-B, 32x3#96Acc@1: 80.2Acc@5: 94.4
action-classification-on-kinetics-400MViT-B, 16x4#118Acc@1: 78.4Acc@5: 93.5
action-classification-on-kinetics-400MViT-S#145Acc@1: 76Acc@5: 92.1
image-classification-on-imagenetMViT-B-24#304Top 1 Accuracy: 84.8%Number of params: 72.9MGFLOPs: 32.7
image-classification-on-imagenetMViT-B-16#508Top 1 Accuracy: 83.0%Number of params: 37MGFLOPs: 7.8