A vehicle re-identification method fused with multi-scale and multi-granularity features

Abstract

Listen

Translate article

Vehicle re-identification is to recognize the same vehicle through vehicle images from different perspectives shot by multiple cameras. At present, the most of vehicle re-identification methods uses convolutional neural network to extract image features but lack the subtle feature discrimination ability. Therefore, we propose a novel vehicle re-identification methods called MSMG Net (Multi-Scale and Multi-Granularity) to solve this problem. We use ResNet-50 as the backbone network, and we design the multi-scale method to extract features of multi-granularity, which is from convolutional layer 3 to convolutional layer 5. For convolutional layer 3, we make the feature map to be divide global feature and local feature of three parts which dimension is 256. For convolutional layer 4, we make the feature map as layer 3 but has only one difference is two parts of local feature. For convolutional layer 5, we extract the global feature from global pooling which dimension is 256. Finally, we fuse these above eight features and get the 2048-dimension to compute the triplet loss and cross-entropy loss. Experiments on the Veri-776 dataset show that the proposed method achieves better performance, and good retrieval results have been achieved on the mAP(74.67%) and Rank-1(94.87%) evaluation index.