RTD-Net (ICCV 2021)

This repo holds the codes of paper: "Relaxed Transformer Decoders for Direct Action Proposal Generation", accepted in ICCV 2021.

News

[2022.4.4] We release codes, checkpoint and features on ActivityNet-1.3.
[2021.8.17] We release codes, checkpoint and features on THUMOS14.

Overview

This paper presents a simple and end-to-end learnable framework (RTD-Net) for direct action proposal generation, by re-purposing a Transformer-alike architecture. Thanks to the parallel decoding of multiple proposals with explicit context modeling, our RTD-Net outperforms the previous state-of-the-art methods in temporal action proposal generation task on THUMOS14 and also yields a superior performance for action detection on this dataset. In addition, free of NMS post-processing, our detection pipeline is more efficient than previous methods.

Dependencies

• Python 3.7 or higher
• PyTorch 1.6 or higher
• Torchvision
• Numpy 1.19.2

Data Preparation

To reproduce the results in THUMOS14 without further changes:

1. Download the data from GoogleDrive.

2. Place I3D_features and TEM_scores into the folder data.

Checkpoint

Dataset	AR@50	AR@100	AR@200	AR@500	checkpoint
THUMOS14	41.52	49.33	56.41	62.91	link

Training

Use train.sh to train RTD-Net.

# First stage

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=11323 --use_env main.py --window_size 100 --batch_size 32 --stage 1 --num_queries 32 --point_prob_normalize

# Second stage for relaxation mechanism

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=11324 --use_env main.py --window_size 100 --batch_size 32 --lr 1e-5 --stage 2 --epochs 10 --lr_drop 5 --num_queries 32 --point_prob_normalize --load outputs/checkpoint_best_sum_ar.pth

# Third stage for completeness head

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=11325 --use_env main.py --window_size 100 --batch_size 32 --lr 1e-4 --stage 3 --epochs 20 --num_queries 32 --point_prob_normalize --load outputs/checkpoint_best_sum_ar.pth

Testing

Inference with test.sh.

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=11325 --use_env main.py --window_size 100 --batch_size 32 --lr 1e-4 --stage 3 --epochs 20 --num_queries 32 --point_prob_normalize --eval --resume outputs/checkpoint_best_sum_ar.pth

References

We especially thank the contributors of the BSN, G-TAD and DETR for providing helpful code.

Citations

If you think our work is helpful, please feel free to cite our paper.

@InProceedings{Tan_2021_RTD,
    author    = {Tan, Jing and Tang, Jiaqi and Wang, Limin and Wu, Gangshan},
    title     = {Relaxed Transformer Decoders for Direct Action Proposal Generation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {13526-13535}
}

Contact

For any question, please file an issue or contact

Jing Tan: jtan@smail.nju.edu.cn
Jiaqi Tang: jqtang@smail.nju.edu.cn