Vehicle re-identification based on keypoint segmentation of original image

In the earlier days, part segmentation methods for vehicle re-id were based on segmenting the feature map of the last convolutional layer. However, by calculating the receptive field, we can see that the size of the receptive field of each point in the feature map of the last convolutional layer exceeds that of the original input image. Therefore, it is very difficult to segment vehicle parts according to feature map. In order to overcome such difficulty, we propose a vehicle re-identification method based on keypoint segmentation of the original image (KSOI). We segment the original image into two parts, which are termed as the window part (upper part) and the below-window part (lower part). Then we use three branches to extract the features of the window part image, the below-window part image and the original vehicle image respectively, which will be fused in the inference stage. In order to achieve accurate segmentation, we label the orientations for all the images in the training set and the keypoint coordinates of the two bottom vertices of the rectangle bounding boxes of the visible windows for some images in the training set. We then train an orientation extraction network and a keypoint detection network to obtain the visible keypoints, and segment the original image with the coordinates of visible keypoints. The experimental results of the proposed KSOI reach the state-of-the-art level.