CSMGAN
Code for ACM MM2020 paper
Jointly Cross- and Self-Modal Graph Attention Network for Query-Based Moment Localization
[Paper]
Main Results
Activity Caption
R@1, IoU=0.3 |
R@1, IoU=0.5 |
R@1, IoU=0.7 |
R@5, IoU=0.3 |
R@5, IoU=0.5 |
R@5, IoU=0.7 |
68.52 |
49.11 |
29.15 |
87.68 |
77.43 |
59.63 |
TACoS
R@1, IoU=0.1 |
R@1, IoU=0.3 |
R@1, IoU=0.5 |
R@5, IoU=0.1 |
R@5, IoU=0.3 |
R@5, IoU=0.5 |
42.74 |
33.90 |
27.09 |
68.97 |
53.98 |
41.22 |
Charades-STA
R@1, IoU=0.5 |
R@1, IoU=0.7 |
R@5, IoU=0.5 |
R@5, IoU=0.7 |
60.04 |
37.34 |
89.01 |
61.85 |
DiDeMo
R@1, IoU=0.5 |
R@1, IoU=0.7 |
R@5, IoU=0.5 |
R@5, IoU=0.7 |
29.44 |
19.16 |
70.77 |
41.61 |
Prerequisites
- Python 3.6
- Pytorch >= 0.4.0
Preparation
Evaluation
$ python main.py --word2vec-path /yourpath/glove_model.bin --dataset ActivityNet --feature-path /yourpath/ActivityCaptions/ActivityC3D --train-data data/activity/train_data_gcn.json --val-data data/activity/val_data_gcn.json --test-data data/activity/test_data_gcn.json --max-num-epochs 20 --dropout 0.2 --warmup-updates 300 --warmup-init-lr 1e-06 --lr 8e-4 --num-heads 4 --num-gcn-layers 2 --num-attn-layers 2 --weight-decay 1e-7 --evaluate --model-load-path ./models_activity/model_6852
$ python main.py --word2vec-path /yourpath/glove_model.bin --dataset TACOS --feature-path /yourpath/TACOS/TACOS --train-data data/tacos/TACOS_train_gcn.json --val-data data/tacos/TACOS_val_gcn.json --test-data data/tacos/TACOS_test_gcn.json --max-num-epochs 40 --dropout 0.2 --warmup-updates 300 --warmup-init-lr 1e-07 --lr 4e-4 --num-heads 4 --num-gcn-layers 2 --num-attn-layers 2 --weight-decay 1e-8 --evaluate --model-saved-path models_tacos --batch-size 64 --model-load-path ./models_tacos/model_4274
Citation
If you use this code please cite:
@inproceedings{liu2020jointly,
title={Jointly Cross- and Self-Modal Graph Attention Network for Query-Based Moment Localization},
author={Liu, Daizong and Qu, Xiaoye and Liu, Xiaoyang and Dong, Jianfeng and Zhou, Pan and Xu, Zichuan},
booktitle={Proceedings of the 28th ACM International Conference on Multimedia (MM’20)},
year={2020}
}