Abstract: In real-world video surveillance applications, person re-identification (ReID) suffers from the effects of occlusions and detection errors. Despite recent advances, occlusions continue to corrupt the features extracted by state-of-art CNN backbones and thereby deteriorate the accuracy of ReID systems. To address this issue, methods in the literature rely on an additional costly process, such as pose estimation, where pose maps provide supervision to focus on visible parts of occluded regions. In contrast, we introduce a Holistic Guidance (HG) method that relies on holistic (or non-occluded) data and its distribution in the dissimilarity space to train the CNN backbone on an occluded dataset. This method is motivated by our empirical study, where the distribution of pairwise between-class and within-class matching distances (Distribution of Class Distances or DCDs) between images has considerable overlap in occluded datasets compared to holistic datasets. Hence, our HG method employs this discrepancy in DCDs of both datasets for joint learning of a student-teacher model to produce an attention map that focuses primarily on visible regions of the occluded images. In particular, features extracted from both datasets are jointly learned using the student model to produce an attention map that allows dissociating visible regions from occluded ones. Additionally, a joint generative-discriminative CNN backbone is trained using a denoising autoencoder such that the system can self-recover from occlusions. Extensive experiments on several challenging public datasets indicate that the proposed approach can outperform state-of-the-art methods on both occluded and holistic datasets.