[ICMR21] Few-Shot Action Localization without Knowing Boundaries

Link:

[Arxiv] [Project] [Presentation]

If you find this helps your research, please cite:

@article{Xie2021FewShotAL,
  title={Few-Shot Action Localization without Knowing Boundaries},
  author={Tingting Xie and Christos Tzelepis and Fan Fu and Ioannis Patras},
  journal={Proceedings of the 2021 International Conference on Multimedia Retrieval},
  year={2021}
}

---

Contents

• Install

• Download

• Training

• Use reference models for evaluation

• Without Learning

• Related project

• Contact

---

Install

git clone https://github.com/June01/WFSAL-icmr21
cd WFSAL-icmr21
pip install -r requirements.txt

Download

Please first create a data dir and then put all the features and annotations under it.

mkdir data
cd data

The feature and annotations used in this paper are originated from wtalc. The features for Thumos14 and ActivityNet1.2 dataset can be downloaded here, while annotations can be found in the original repo.

Training (5-way 1-shot)

For thumos 14,

python main.py --split='cvpr18' --encoder --num_in=4 --tsm='ip' --sample_num_per_class=1 --batch_num_per_class=5 

For ActivityNet1.2:

python main.py --split='cvpr18' --dataset='ActivityNet1.2' --num_in=4 --encoder --tsm='ip' --sample_num_per_class=1

For evaluation,

python main.py --split='cvpr18' --dataset=dataset --num_in=4 --encoder --tsm=ip --sample_num_per_class=1 --mode=testing --load=/path/to/model

Note, we report the median of 10 repetitions.

Without Learning

python eval_non_learning.py

Related project

• wtalc-pytorch

• 3c-net

• DGAM-Weakly-Supervised-Action-Localization

Contact

We would like thank Dr. Hongtao Yang for his useful discussions around the tasks discussed in this paper as well as the evaluation metrics used. For any question, please file an issue or contact

[email protected]