[CVPR 2024 Accepted] Task-Driven Exploration: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection
Task-Driven Exploration: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection
This code repo implements TaskWeave in CVPR 2024, the first attempt to explore the task-driven paradigm for joint Moment Retrieval and Highlight Detection. In this paper, we present the first task-driven top-down framework, named TaskWeave. We introduce a task-decoupled unit to capture task-specific and common representations. To further investigate the interactions between these two tasks, we propose an inter-task feedback mechanism. It transforms the results of one task into guiding masks to assist the other task. Lastly, different from existing methods, we present a task-dependent joint loss function to optimize the model. As far as we are aware, this is the first framework to address this joint task from the task-centric perspective. Comprehensive experiments and in-depth ablation studies on QVHighlights, TVSum, and Charades-STA datasets corroborate the effectiveness and flexibility of the proposed framework.
Please refer to MomentDETR for more details. Please refer to UMT for more details. Please refer to QD-DETR for more details.
- Train(Take
QVHighlights
as an example)
bash taskweave/scripts/train.sh
bash taskweave/scripts/train_audio.sh
- Evaluation (Take
QVHighlights
as an example)
bash taskweave/scripts/inference.sh results/{direc}/model_best.ckpt 'val'
bash taskweave/scripts/inference.sh results/{direc}/model_best.ckpt 'test'
If you are using our code, please consider citing the following paper.
@inproceedings{yang2024taskweave,
title={Task-Driven Exploration: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection},
author={Yang, Jin and Wei, Ping and Li, Huan and Ren, Ziyang}
booktitle={CVPR},
year={2024}
}