I3D, I3D Nonlocal

August 2, 2021 · View on GitHub

Use following model configs.

model_configs/i3d_resnet50.py
model_configs/i3dnonlocal_resnet50.py

Kinetics

architecture	depth	pretrain	frame length x sample rate	top1	top5	model	config
I3D	R50	-	8 x 8	73.5	90.8	`link`	Kinetics/c2/I3D_8x8_R50
I3D NLN	R50	-	8 x 8	74.0	91.1	`link`	Kinetics/c2/I3D_NLN_8x8_R50

Reference: facebookresearch/video-nonlocal-net

_script	_{input frames}	_{freeze bn?}	_{3D conv?}	_non-local?	_top1	_{in paper}	_top5	_model	_logs
_{run_i3d_baseline_400k_32f.sh}	32	-	Yes	-	73.6	_73.3	90.8	`link`	`link`
_{run_i3d_nlnet_400k_32f.sh}	32	-	Yes	Yes	74.9	_74.9	91.6	`link`	`link`

This model is trained with PyVideoAI.
Top1/5 accuracy is calculated using 1 spatial centre crop and 5 temporal crops.
dataset_configs/hmdb.py
model_configs/i3d_resnet50.py
exp_configs/hmdb/i3d_resnet50-crop224_lr0001_batch8_8x8_largejit_plateau_1scrop5tcrop_split1.py

architecture	Pretrain	frame length x sampling stride	Top1 (highest/last)	Top5 (highest/last)	config, log	model (last)	TensorBoard
I3D-ResNet50	Kinetics	8 x 8	73.20 / 72.94	94.05 / 94.05	`link`	`link`	`link`

ImageNet

Reference: facebookresearch/SlowFast

architecture	depth	Top1	Top5	model
ResNet	R50	23.6	6.8	`link`

Kinetics

HMDB-51

ImageNet