# Evidential Deep Learning for Open Set Action Recognition

**Authors:** Wentao Bao, Qi Yu, Yu Kong  
**arXiv:** [2107.10161](https://arxiv.org/abs/2107.10161)  
**Date:** 2021-07-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.10161v1/

## Key points

- The paper addresses the challenge of recognizing human actions in a real-world scenario
- It introduces a novel method called Deep Evidential Action Recognition (DEAR) for open set action recognition
- DEAR can recognize actions in an open testing set and reject unknown actions
- The approach formulates the action recognition task from the perspective of evidential deep learning (EDL)
- A model calibration technique is introduced to regularize EDL training
- A plug-and-play module is proposed to debias video representations through contrastive learning
- Experimental results show that DEAR consistently improves performance across multiple action recognition models and benchmarks
- DEAR effectively addresses challenges related to out-of-distribution human actions and static bias in video representations
- The authors plan to make their codes and pre-trained weights available upon acceptance of the paper.

## Summary

The paper titled "Evidential Deep Learning for Open Set Action Recognition" addresses the challenge of recognizing human actions in a real-world scenario, where the actions performed by individuals may differ from those present in the training data. This necessitates a model that can not only identify known actions but also reject unknown ones. Unlike image data, recognizing video actions in an open-set setting is more difficult due to the uncertain temporal dynamics and static bias associated with human actions. To tackle this problem, the authors propose a novel method called Deep Evidential Action Recognition (DEAR) that can recognize actions in an open testing set. The approach formulates the action recognition task from the perspective of evidential deep learning (EDL) and introduces a model calibration technique to regularize EDL training. Additionally, to mitigate the static bias inherent in video representations, a plug-and-play module is proposed to debias the learned representation through contrastive learning. Experimental results demonstrate that the DEAR method consistently improves performance across multiple mainstream action recognition models and benchmarks. In summary, this paper presents a novel approach for open set action recognition using evidential deep learning. The proposed DEAR method effectively addresses challenges related to out-of-distribution human actions and static bias in video representations, leading to improved performance compared to existing approaches. The authors plan to make their codes and pre-trained weights available upon acceptance of the paper.