#You Can Only Watch the Past：Track Attention Network for Online Spatio-Temporal Action Detection

Overview of TAN

Requirements

• We recommend you to use Anaconda to create a conda environment:

conda create -n tan python=3.6

• Then, activate the environment:

conda activate tan

• Requirements:

pip install -r requirements.txt 

Visualization

Dataset

UCF101-24:

You can download UCF24 from the following links:

• Google drive

Link: https://drive.google.com/file/d/1Dwh90pRi7uGkH5qLRjQIFiEmMJrAog5J/view?usp=sharing

• BaiduYun Disk

Link: https://pan.baidu.com/s/11GZvbV0oAzBhNDVKXsVGKg

Password: hmu6

Train TAN

• UCF101-24

For example:

python -m torch.distributed.launch --nnodes 1 --nproc_per_node 2 train.py --cuda -d ucf24 --data_root /data1/su/datasets/UCF24-YOWO/ -bs 32 -tbs 16 -K 16 -accu 8 -v yowo_v3_large --max_epoch 7 --lr_epoch 2 3 4 5 --eval -ct 0.05 --distributed --sybn (--resume weights/ucf24/yowo_v3_large/yowo_v3_large_epoch_1.pth) (--untrimmed_trainning) (--track_mode)

Test TAN

• Frame-mAP：

python eval.py --cuda -d ucf24 --data_root /data1/su/datasets/UCF24-YOWO/ -tbs 16 -v tan_large --weight weights/ucf24/tan_large/tan_large_epoch_0.pth -ct 0.05 --cal_frame_mAP

• Video mAP：

python eval.py --cuda -d ucf24 --data_root /data1/su/datasets/UCF24-YOWO/ -tbs 16 -v tan_large --weight weights/ucf24/tan_large/tan_large_epoch_0.pth -ct 0.05 --cal_video_mAP --link_method viterbi

Note

Model weights and detection results can be downloaded from the cloud drive link below. Link: https://pan.baidu.com/s/1j4890Y6rtzycWG_jQeyd-Q?pwd=m5xd password: m5xd

Currently, only part of the code is released. The complete code will be released after the paper is accepted.