[ICCV 2021] Relaxed Transformer Decoders for Direct Action Proposal Generation

Last update: Nov 30, 2022

Overview

RTD-Net (ICCV 2021)

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

News

[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:

Download the data from GoogleDrive.
Place I3D_features and TEM_scores into the folder data.

Checkpoint

Dataset	[email protected]	[email protected]	[email protected]	[email protected]	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: [email protected]
Jiaqi Tang: [email protected]

[ICCV 2021] Relaxed Transformer Decoders for Direct Action Proposal Generation

Related tags

Overview

RTD-Net (ICCV 2021)

News

Overview

Dependencies

Data Preparation

Checkpoint

Training

Testing

References

Citations

Contact

Owner

Multimedia Computing Group, Nanjing University

The backbone CSPDarkNet of YOLOX.

CONetV2: Efficient Auto-Channel Size Optimization for CNNs

PyTorch code for Composing Partial Differential Equations with Physics-Aware Neural Networks

Twin-deep neural network for semi-supervised learning of materials properties

Hardware accelerated, batchable and differentiable optimizers in JAX.

Repository for scripts and notebooks from the book: Programming PyTorch for Deep Learning

Plato: A New Framework for Federated Learning Research

MassiveSumm: a very large-scale, very multilingual, news summarisation dataset

Instance-wise Feature Importance in Time (FIT)

Semantic Segmentation for Real Point Cloud Scenes via Bilateral Augmentation and Adaptive Fusion (CVPR 2021)

Understanding and Improving Encoder Layer Fusion in Sequence-to-Sequence Learning (ICLR 2021)

meProp: Sparsified Back Propagation for Accelerated Deep Learning

PyTorch Kafka Dataset: A definition of a dataset to get training data from Kafka.

An alarm clock coded in Python 3 with Tkinter

The spiritual successor to knockknock for PyTorch Lightning, get notified when your training ends

Learning with Noisy Labels via Sparse Regularization, ICCV2021

[AAAI2021] The source code for our paper 《Enhancing Unsupervised Video Representation Learning by Decoupling the Scene and the Motion》.

Source code of our TTH paper: Targeted Trojan-Horse Attacks on Language-based Image Retrieval.

TrackTech: Real-time tracking of subjects and objects on multiple cameras

Code for CVPR 2018 paper --- Texture Mapping for 3D Reconstruction with RGB-D Sensor