Sparse Label Smoothing for Semi-supervised Person Re-Identification
In this paper, we propose a semi-supervised framework to address the over-smoothness problem found in current regularization methods. We carefully propose to derive a regularization method by constructing clusters of similar images. We propose Sparse Label Smoothing Regularization (SLSR) which consist of three steps. First, we train a CNN to learn discriminative patterns from labeled data. For each image, we extract the feature map from the last convolution layer and directly apply k-means clustering algorithm on the feature. Secondly, we train a GAN model for feature representation learning and generate sample images for each cluster. Each generated sample is assigned a label using our regularization method. Thirdly, we define a new objective function and fine-tuned two baseline models ResNet and DenseNet. Extensive experiments on four large-scale datasets Market-1501, CUHK03, DukeMTMC-ReID, and VIPeR show that our regularization method significantly improves the Re-ID accuracy compared to existing semi-supervised methods. On Market-1501 dataset, for instance, rank-1 accuracy is improved from 87.29 from 90.05 https://github.com/jpainam/SLS_ReID
READ FULL TEXT