Official repository of ACM MM 2021 paper "Few-Shot Fine-Grained Action Recognition via Bidirectional Attention and Contrastive Meta-Learning"

Fine-grained action recognition is attracting increasing attention due to the emerging demand of specific action understanding in real-world applications, whereas the data of rare fine-grained categories is very limited. Therefore, we propose the few-shot fine-grained action recognition problem, aiming to recognize novel fine-grained actions with only few samples given for each class. Although progress has been made in coarse-grained actions, existing few-shot recognition methods encounter two issues handling fine-grained actions: the inability to capture subtle action details and the inadequacy in learning from data with low inter-class variance. To tackle the first issue, a human vision inspired bidirectional attention module (BAM) is proposed. Combining top-down task-driven signals with bottom-up salient stimuli, BAM captures subtle action details by accurately highlighting informative spatio-temporal regions. To address the second issue, we introduce contrastive meta-learning (CML). Compared with the widely adopted ProtoNet-based method, CML generates more discriminative video representations for low inter-class variance data, since it makes full use of potential contrastive pairs in each training episode. Furthermore, to fairly compare different models, we establish specific benchmark protocols on two large-scale fine-grained action recognition datasets. Extensive experiments show that our method consistently achieves state-of-the-art performance across evaluated tasks.

Overview of our framework Visualization of attention maps generated by our bidirectional attention module (BAM) Comparision of video representations generated by ProtoNet and the proposed contrastive meta-learning (CML)

The evaluations are performed on two publicly released fine-grained action recognition datasets: FineGym and HAA500.

Please visit the official site of FineGym to download data and extract frames for video clips with Gym99 and Gym288 annotations.

Please visit the official site of HAA500 and download the 1.1 version of the dataset. Directly extract frames for each video since they are already trimmed.

You can find the split files of our few-shot fine-grained action recognition benchmarks here. The specific training/test protocols of Gym99, Gym288 and HAA500 are introduced in Sec. 4.1 of our paper.

For each benchmark, "test_class.txt" and "train_class.txt" contain the labels of training/test categories. "train_split.txt" and "test_split.txt" contain the annotations for dataloaders in TPN style, i.e. each line is a record of a video sample in the form of: directory path of a video, total frame number, groundtruth label.

If you find this repository useful, please cite:

@inproceedings{wang2021few-shot,
  title={Few-shot fine-grained action recognition via bidirectional attention and contrastive meta-learning},
  author={Wang, Jiahao and Wang, Yunhong and Liu, Sheng and Li, Annan},
  booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
  year={2021}
}