Dynamic locally connected layer for person re-identification

Person re-identification is a challenging task due to its large variations on pedestrian pose, camera view, lighting and background. To solve pedestrian misalignment problem, most of the existing works assume that the pedestrian images are horizontally aligned so that the extracted features can be compared correspondingly. However, such assumption is not necessarily true in reality because the pedestrians may be misaligned vertically. To address the misalignment problem, we propose a dynamic locally connected (DLC) layer based on convolutional neural network (CNN). We use human parsing tool to get parsing results of pedestrian images, then map the results to the last feature map of our CNN. By doing this, proposed model is able to locate the human body parts dynamically within DLC layer, thus leads to a more accurate matching on local features. Furthermore, we adopt pre-training with two-step fine-tuning strategy on the small person re-identification datasets, which again boost the model performance. According to the experiments, proposed model achieves competitive results among the state-of-the-art models on four popular person re-identification datasets.