BasicTAD

Our paper is available in BasicTAD

News

[2023.5.4] New code is updated.
[2023.4.13] Our BasicTAD has been published in CVIU. New code will be updated soon.
[2022.11.14] We upload pretrained backbone for users to train their own models.
[2022.11.2] Some issues have been fixed. More complete code is on the way.
[2022.9.1] We update README.md, and release codes, checkpoint on THUMOS14.

alt

Overview

This paper is empirical study on end-to-end TAD pipeline. Here we release our code here for further study of the TAD task. We hope to contribute to the development of the TAD community.

Environment preparation

Create environment

conda create -n basictad python=3.8

Activate environment

conda activate basictad

Install pytorch (take cuda==10.2 as example)

conda install pytorch==1.8.1 torchvision==0.9.1 torchaudio==0.8.1 cudatoolkit=10.2 -c pytorch

Install mmcv (take mmcv-full==1.4 as example)

pip install mmcv-full==1.4 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.8.0/index.html

Clone the basictad repository.

git clone https://github.com/MCG-NJU/BasicTAD.git
cd BasicTAD

Install basictad and other dependencies

pip install -r requirements/build.txt
pip install -v -e .

Data preparation

Thumos14

1: Download datasets

cd ${basictad_root}/data/thumos14
bash download.sh

2: Extract frames

cd ${basictad_root}/data/thumos14
#3fps:
bash extract_frames.sh videos/val frames_3fps/validation -vf fps=3 %05d.jpg
bash extract_frames.sh videos/test frames_3fps/test -vf fps=3 %05d.jpg
#6fps:
bash extract_frames.sh videos/val frames_6fps/validation -vf fps=6 %05d.jpg
bash extract_frames.sh videos/test frames_6fps/test -vf fps=6 %05d.jpg

3: We upload two files in this link(key:cjsk) as pretained backbone SlowOnly. Put SLOW_8x8_R50.pyth into ～/.cache/toch/hub/checkpoints. Unzip facebookresearch_pytorchvideo_master.zip into ～/.cache/toch/hub

Train and test

cd ${basictad_root}
#anchor-free-3fps
bash tools/thumos/train_and_test_thumos_anchor_free_3fps.sh
#anchor-free-6fps
bash tools/thumos/train_and_test_thumos_anchor_free_6fps.sh
#anchor-based-3fps
bash tools/thumos/train_and_test_thumos_anchor_based_3fps.sh
#anchor-based-6fps
bash tools/thumos/train_and_test_thumos_anchor_based_6fps.sh

Checkpoint

Method	mAP@0.3	mAP@0.4	mAP@0.5	mAP@0.6	mAP@0.7	Avg	checkpoint
anchor_based_6fps	72.3	68.4	62.0	52.4	37.0	58.4	link(key:z509)
anchor_free_6fps	75.1	70.2	63.0	50.6	38.7	59.5	link(key:kkn3)

How to use Checkpoints above

cd ${basictad_root}
# anchor_based_6fps
CUDA_VISIBLE_DEVICES=0 python tools/thumos/test_ab.py configs/trainval/basictad/thumos14/basictad_slowonly_e700_thumos14_rgb_192win_anchor_based.py anchor_based-6fps/epoch_300_epoch.pth
# anchor_free_6fps
CUDA_VISIBLE_DEVICES=0 python tools/thumos/test_af.py --framerate 6 configs/trainval/basictad/thumos14/basictad_slowonly_e700_thumos14_rgb_192win_anchor_free.py anchor_free-6fps/epoch_600_epoch.pth

Credits

We especially thank the contributors of the DaoTAD for providing helpful code.

MCG-NJU / BasicTAD

readme