Awesome

SHINE: Saliency-aware HIerarchical NEgative Ranking for Compositional Temporal Grounding -- ECCV2024

This is the implementation for the paper "SHINE: Saliency-aware HIerarchical NEgative Ranking for Compositional Temporal Grounding" (ECCV 2024): paper, ArXiv version.

by

Zixu Cheng*¹, Yujiang Pu*², Shaogang Gong¹, Parisa Kordjamshidi², Yu Kong²

¹Queen Mary University of London, ² Michigan State University (* Equal Contribution)

Prerequisites

<b>0. Clone this repo</b>

<b>1. Prepare datasets</b>

<b>Charades-CG</b> : Download I3D feature files for Charades-CG dataset from VSLNet.

<b>ActivityNet-CG</b> : Download C3D feature files for ActivityNet-CG dataset from MS-2D-TAN.

Text Features : We provide our hierarchical negative query features here. (DropBox)

<b>2. Install dependencies.</b>

conda create -n shine python=3.10
conda activate shine
cd SHINE
pip install -r requirements.txt

Train

Charades-CG

1. Add your data and feature path in shine/scripts/train_charades_cg.sh

######## setup video+text features
feat_root= # path/to/your/anet/features

2. Run the script

bash shine/scripts/train_charades_cg.sh

ActivityNet-CG

1. Add your data and feature path in shine/scripts/train_anet_cg.sh

######## setup video+text features
feat_root= # path/to/your/anet/features

2. Run the script

bash shine/scripts/train_anet_cg.sh --clip_length 1 --saliency_margin 1.0  --max_es_cnt 10 --max_q_l 50 --enc_layers 3 --dec_layers 3 --lr 0.00013 --use_saliency_loss

Evaluation

# Evaluate Charades-CG
bash shine/scripts/inference_charades.sh path/to/your/ckpt 'val'
# Evaluate ActivityNet-CG
bash shine/scripts/inference_anet.sh path/to/your/ckpt 'val'

We also provide our checkpoints here. (DropBox)

Contributors and Contact

If there are any questions, feel free to contact the authors: Zixu Cheng (zixu.cheng@qmul.ac.uk), and Yujiang Pu (puyujian@msu.edu).

Acknowledgment

Our implementations are based on Moment-DETR and QD-DETR. We thank the authors for their awesome open-source contributions.

LICENSE

The annotation files are transformed from VISA and many parts of the implementations are borrowed from Moment-DETR and QD-DETR. Following, Our codes are also under the MIT license.