E2E-TAD

This repository holds the code for the following paper:

An Empirical Study of End-to-end Temporal Action Detection
Xiaolong Liu, Song Bai, Xiang Bai
CVPR 2022.

This paper presents an empirical study of end-to-end temporal action detection (TAD). It

• reveals the benefit of end-to-end training. We observe up to 11% performance improvement.
• studies the effect of a series of design choices in end-to-end TAD, including detection head, video encoder, spatial and temporal resolution of videos, frame sampling manner and multi-scale feature fusion.
• establishes a baseline detector. Built upon SlowFast and TadTR, it outperforms previous SOTA methods such as MUSES and AFSD with more than 4x faster speed, using only RGB modality. It can process 5076 frames per second on a single TITAN Xp GPU.
  

We hope that E2E-TAD can accelerate the research and applications of end-to-end temporal action detection.

This code is an extended version of TadTR. It supports both video inputs and video feature inputs. If you want to run with video features, please refer to the instruction here.

TODOs

• Inference code

• Training code

Updates

Feb 19, 2023: Fix a bug in loss calculation.

Aug 14, 2022: The inference code for THUMOS14 dataset is released.

Apr, 2022: This repo is online.

Mar, 2022: Our paper is accepted by CVPR 2022.

Results

THUMOS14

Config	Encoder	Head	SR, TR	AmAP	GPU Mem.	Training Speed
link	SlowFast R50 4x16	TadTR	96, 10FPS	54.2	7.6G	17 min/epoch

ActivityNet

Config	Encoder	Head	SR, TR	AmAP	GPU Mem.	Train Speed
link	SlowFast R50 4x16	TadTR	96, 384	35.10	11G	62 (30)* min/epoch
link	TSM R50	TadTR	96, 96	34.14	10G	30 (19) min/epoch

* The values in the brackets are measured on RTX 3090, others on TITAN Xp.

SR: spatial (image) resolution. TR: temporal resolution, measured by the sampling frame rate on THUMOS14 and the number of sampled frames per video on ActivityNet.

0.Install

Requirements

• Linux or Windows. Better with SSD, because end-to-end has high IO demand.

• Python>=3.7

• CUDA>=9.2, GCC>=5.4

• PyTorch>=1.5.1, torchvision>=0.6.1 (following instructions here)

• Other requirements

  pip install -r requirements.txt

Compiling CUDA extensions

The RoIAlign operator is implemented with CUDA extension.

cd model/ops;

# If you have multiple installations of CUDA Toolkits, you'd better add a prefix
# CUDA_HOME=<your_cuda_toolkit_path> to specify the correct version. 
python setup.py build_ext --inplace

Run a quick test

python demo.py --cfg configs/thumos14_e2e_slowfast_tadtr.yml

1.Data Preparation

THUMOS14

Download video frames and annotation files from [BaiduDrive](code: adTR) or [OneDrive].

Put all the following data under data/thumos14 directory.

• Annotations: The annotations of action instances and the meta information of video files. Both are in JSON format (th14_annotations_with_fps_duration.json and th14_img10fps_info.json). The meta information file records the FPS and number of frame of each video. You can generated it by yourself using the script tools/prepare_data.py.
• Video Frames: We provide video frames extracted at 10fps. Extract the archive using tar -xf thumos14_img10fps.tar.
• Pre-trained Reference Model: Our pretrained model that uses SlowFast R50 encoder thumos14_e2e_slowfast_tadtr_reference.pth. This model corresponds to the config file configs/thumos14_e2e_slowfast_tadtr.yml.

Note that our code does not depend on raw videos. You can download them mannually from the official site if you need them. We also provide a script that can download videos and extract frames. Please refer to tools/extract_frames.py.

2.Testing Pre-trained Models

Run

python main.py --cfg CFG_PATH --eval --resume CKPT_PATH

CFG_PATH is the path to the YAML-format config file that defines the experimental setting. For example, configs/thumos14_e2e_slowfast_tadtr.yml. CKPT_PATH is the path of the pre-trained model. Alternatively, you can execute the Shell script bash scripts/test_reference_models_e2e.sh thumos14 for simplity.

3.Training by Yourself

To be done. We are still checking the codebase. Plan to add official support of training in the next week later.

Using Our Code on Your Dataset

Please refer to docs/1_train_on_your_dataset.md.

Acknowledgement

The code is based on our previous project TadTR. We also borrow some code from DETR, Deformable DETR, G-TAD, TSM and MMAction2. Thanks for their great works.

Citation

@inproceedings{liu2022an,
  title={An Empirical Study of End-to-end Temporal Action Detection},
  author={Liu, Xiaolong and Bai, Song and Bai, Xiang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={20010-20019},
  year={2022}
}

Related Projects

• TadTR: an early version of this project. TadTR is an efficient and flexible Transformer network for temporal action detection.