DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer

This repository contains the official implementation of the paper DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer.

[paper] [CVPR paper]

Installation

NMS

cd util
python setup.py install --user # build NMS
cd ..

Temporal Deformable Attention

cd models/digit/ops
python setup.py build install
cd ../../..

Prepare Dataset

We follow ActionFormer repository and Video Mamba Suite for preparing datasets including THUMOS14, ActivityNet v1.3, and HACS-Segment.

Use scripts/make_feature_info.py to generate feature information for each dataset. THUMOS14 is already prepared in the repository.

Training

To train the DiGIT model on the THUMOS14 dataset, execute the following command:

python main.py --c config/digit/internvideo2/thumos14.py --output_dir logs/thumos14

Evaluation

To evaluate the trained model and obtain performance metrics, use the following command structure:

python main.py --eval --c config/digit/internvideo2/thumos14.py --output_dir logs/thumos14

Citation

if you find our work helpful, please consider citing our paper:

@InProceedings{Kim_2025_CVPR,
    author    = {Kim, Ho-Joong and Lee, Yearang and Hong, Jung-Ho and Lee, Seong-Whan},
    title     = {DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {24286-24296}
}