action_detection

Action Detection

Installation

Please follow the installation instructions in INSTALL.

⚠️ Note that you have to set ava datapath DATA_DIR in alphaction/config/paths_catalog.py.

Datasets

You can find the dataset instructions in DATASET.

Model ZOO

Model	Setting	#Frame	mAP	Model	Shell	Log
UMT-B/16	K710 PT+FT	8	33.5	ckpt	run.sh	log
UMT-L/16	K710 PT+FT	8	39.8	ckpt	run.sh	log

Finetuning

For finetuning, you can simply run the pretraining scripts in exp as follows:

bash ./exp/b16_ptk710_ftk710_ftava_f8_res224.sh

⚠️ Notes:

1. Chage MODEL_PATH to your model path.
2. For ViT-L, we set --close_amp to close amp for stable finetuning.
3. Since we pretrain the models on Kinetics with sparse sampling, the frame span should be similar for better performance. Thus we set --sparse to use a large frame span 300, which is equal to the average frame number of Kinetics clips.