Unsupervised Group Re-identification via Transferred Single and Couple Representation Learning

http://arxiv.org/abs/1905.04854

Abstract: Group re-identification (G-ReID) is a less-studied task. Its challenges not only lie in appearance changes of individuals which have been well-investigated in general person re-identification (ReID), but also derive from group layout and membership changes. So the key task of G-ReID is to learn representations robust to the changes. To address this issue, we design a Transferred Single and Couple Representation Learning Network (TSCN). The merits are two aspects: 1) Due to the lack of training samples, existing methods exploit unsatisfactory hand-crafted features. To obtain the superiority of deep learning models, we treat a group as multiple persons and transfer the labeled ReID dataset to the G-ReID dataset style to learn the single representation. 2) Taking into account neighborhood relationship in a group, we also propose the couple representation, which maintains more discriminative features in some cases. In addition, an unsupervised weight learning method is exploited to adaptively fuse the results of different views together according to the result pattern. Extensive experimental results demonstrate the effectiveness of our approach, outperforming the state-of-the-art method respectively 11.7% CMC-1 on Road Group dataset and 39.0% CMC-1 on DukeMCMT dataset.

Challenge

• Group layout change. People in a group often change their locations under different camera views, as figure shows.

• Group membership change. Also as figure shows, people will dynamically leave or join a group, thereby making the number of persons in the group change over time. To this end, it is not a good choice to treat the group as a whole to extract its global/semi-global features, because the group layout and membership change will alter the visual content of the group.

• Lack of labelled training data. In addition, G-ReID is a new task and lacks annotated image datasets with group IDs, ie, the number of group images is too few to learn robust group representations.

Framework

The proposed architecture consists of two parts: the offline representation learning framework and the online fusion framework. Here, we take DukeMTMC group dataset as the example target-domain dataset. In offline representation learning, we first transfer the style of source-domain dataset (\eg, Market-1501) to that of the target one (\eg, DuketMTMC). We then train the Transferred Single and Couple Network (TSCN) on the transferred labeled dataset. During online funsion, TSCN extracts multiple single and couple features from each group image, and then measure the single and couple representation distance compared to the representations of gallery groups. Through evaluating the ``L'' shape fitness of the curves, we learn the weights for combination. Finally, the fusion results are obtained according to the weights.

Implement

If you want to implement our code:

• You should download dataset at http://min.sjtu.edu.cn/lwydemo/GroupReID.html firstly.
• Use CycleGAN to transfer the style of market1501 into style of target dataset.
• The transferred data should be saved in single/data/market1501 and couple/data/market1501
• Single representation network can be implemented as:
  cd single
python3 main.py -d market1501 --logs-dir logs/market1501
• Couple representation network can be implemented as:
  cd couple
python3 main.py -d market1501 --logs-dir logs/market1501