Kai Jiang’s research while affiliated with Shanghai University of Engineering Science and other places

Person Re-Identification (Re-ID) technology is being developed rapidly due to the successful application of deep convolutional neural networks. However, the prevailing Re-ID models are usually built upon manually design backbones. In this paper, we propose using the TGAS-ReID which is automatically designed convolutional network backbones for Re-ID to substitute the backbones originally designed for classification such as ResNet and VGG. In the Re-ID tasks to search for a cell structure, greedy decisions are made instead of deriving the architecture after comprehensive training. In other words, at each decision epoch, according to the topological order, we first decide the candidates’ pool of the edges to progressively reduce the coupling of the internal nodes of the DAG. An edge is then selected based on edge importance, edge certainty, and selection stability. We then make a greedy optimal choice for the selected edge and prune the relevant parameters. To further improve the backbone’s representation capability of the features, we further introduce the triplet loss with batch hard mining as the retrieval loss. Extensive experiments demonstrate that the searched structure of the backbones reaches a performance level close to the previous work with a 20.8% shorter searching time. The proposed method also prevents the final CNNs network from suffering the well-known performance collapse by avoiding aggregation of the skip-connections.

Person re-Identification(Re-ID) based on deep convolutional neural networks (CNNs) achieves remarkable success with its fast speed. However, prevailing Re-ID models are usually built upon backbones that manually design for classification. In order to automatically design an effective Re-ID architecture, we propose a pedestrian re-identification algorithm based on knowledge distillation, called KDAS-ReID. When the knowledge of the teacher model is transferred to the student model, the importance of knowledge in the teacher model will gradually decrease with the improvement of the performance of the student model. Therefore, instead of applying the distillation loss function directly, we consider using dynamic temperatures during the search stage and training stage. Specifically, we start searching and training at a high temperature and gradually reduce the temperature to 1 so that the student model can better learn from the teacher model through soft targets. Extensive experiments demonstrate that KDAS-ReID performs not only better than other state-of-the-art Re-ID models on three benchmarks, but also better than the teacher model based on the ResNet-50 backbone.