Contextual Active Learning for Person Re- Identification

Active learning has been recently investigated in the field of Person Re-identijication to obtain informative samples for training. However, the current methods incur a fixed annotation cost as they do not explicitly incorporate the re-identification model’ confidence. To this end, we propose a novel human-in-the-loop context aware active learning method that helps the re-identification model improve with progressively collected data while annotating a few but effective samples. In our proposed method, a contex-tual bandit agent is trained to learn a policy to obtain the training samples about which the model is least confident and thus needs annotations. A binary reward is provided to the agent based on the actions and the confidence of the model given the current query image. On an average, our model achieves a boost of 9.13% mAP, 5.64% rank-1 improvement over the baseline and uses 32.3% less anno-tations compared to the previous best active learning approach on DukeMTMC-reID.