Model Zoo

Note

• For all the pretraining and finetuning, we adopt spaese/uniform sampling.
• #Frame $=$ #input_frame $\times$ #crop $\times$ #clip
• #input_frame means how many frames are input for model per inference
• #crop means spatial crops (e.g., 3 for left/right/center)
• #clip means temporal clips (e.g., 4 means repeted sampling four clips with different start indices)

Pretraining

Model	Setting	Model	Shell	Log
UMT-B/16	K710 200e	ckpt	run.sh	log
UMT-L/16	K710 200e	ckpt	run.sh	log

Finetuning

K710

Model	Setting	#Frame	Top-1	Model	Shell	Log
UMT-B/16	K710 PT	8x3x4	81.9	ckpt	run.sh	log
UMT-L/16	K710 PT	8x3x4	86.0	ckpt	run.sh	log

K400

Model	Setting	#Frame	Top-1	Model	Shell	Log
UMT-B/16	K710 PT+FT	8x3x4	87.4	ckpt	run.sh	log
UMT-L/16	K710 PT+FT	8x3x4	90.3	ckpt	run.sh	log
UMT-L/16	K710 PT+FT	16x3x4	90.6	ckpt	run.sh	log

K600

Model	Setting	#Frame	Top-1	Model	Shell	Log
UMT-B/16	K710 PT+FT	8x3x4	87.8	ckpt	run.sh	log
UMT-L/16	K710 PT+FT	8x3x4	90.4	ckpt	run.sh	log
UMT-L/16	K710 PT+FT	16x3x4	90.5	ckpt	run.sh	log

K700

Model	Setting	#Frame	Top-1	Model	Shell	Log
UMT-B/16	K710 PT+FT	8x3x4	78.5	ckpt	run.sh	log
UMT-L/16	K710 PT+FT	8x3x4	83.2	ckpt	run.sh	log
UMT-L/16	K710 PT+FT	16x3x4	83.6	ckpt	run.sh	log

MiT V1

Model	Setting	#Frame	Top-1	Model	Shell	Log
UMT-B/16	K710 PT+FT, K400 FT	8x3x4	44.6	ckpt	run.sh	log
UMT-L/16 384↑	K710 PT+FT, K400 FT	8x3x4	45.5	ckpt	run.sh	log
UMT-L/16	K710 PT+FT, K400 FT	8x3x4	48.0	ckpt	run.sh	log
UMT-L/16 384↑	K710 PT+FT, K400 FT	8x3x4	48.7	ckpt	run.sh	log

SthSth V2

Model	Setting	#Frame	Top-1	Model	Shell	Log
UMT-B/16	K710 PT	8x3x4	70.8	ckpt	run.sh	log
UMT-L/16	K710 PT	8x3x4	74.7	ckpt	run.sh	log

AVA v2.2

See action_detection.