Person Re-Identification Using Attribute Attention Network (AANet)

Person re-identification is an important computer vision task for video surveillance. Key challenges for person re-identification includes occlusion, variation in view angle & lighting condition, etc. To address these issues, we proposed a deep learning model called AANet which encodes global identification feature, local body region features and person attribute representation. We achieve state-of-the-art performances. In addition, we can perform identity retrieval using both image and text queries.
(Published in CVPR2019)

 

Figure 1. Person attribute heatmaps are combined to form the Attribute Attention Map (AAM).

Figure 2. The AAM captures more relevant identity features than the traditional global identity (ID) heatmap and allow our model to encode more discriminative representation.

Figure 3. Overview of the AANet. The feature maps from the ResNet50 backbone are learned concurrently by the Global Feature Network (GFN), Part Feature Network (PFN) and the Attribute Feature Network (AFN).

Figure 4. Examples of how person attributes help in improving the image retrieval accuracy.

Figure 5. Performance comparisons with the state-of-the-arts using DukeMTMC-reID and Market1501 datasets. AANet-50 and AANet-152 denote our proposed model with ResNet50 and ResNet152 backbone respectively. RR denotes Re-Ranking.

For more information, you may contact our professor Yap Kim Hui.