Currently, researchers incline to employ large-scale datasets as benchmarks for pre-training and fine-tuning models on small-scale datasets to achieve superior performance. However, many researchers cannot afford the enormous computational overhead that pre-training entails, and fine-tuning is easy to compromise the generalization ability of models for the target dataset. Therefore, model learning on the small challenging data regime should be given renewed attention, which will benefit many tasks such as person re-identification. To this end, we propose a novel model named “Two-Dimensional Serpentine Network (2D-SNet)”, which is constructed by multiple lightweight and effective “Two-Dimensional Serpentine Blocks (2D-SBlocks)”. The generalization ability of 2D-SNet stems from three points: (a) 2D-SBlock utilizes multi-scale convolution kernels to extract the multi-scale information from images on the small data regime; (b) 2D-SBlock has a serpentine calculation order, which significantly reduces the number of skip connections and can thereby save many computational and storage resources; (c) 2D-SBlock improves the discrimination ability of 2D-SNet via BN-Depthwise Conv or MSA. As experimentally demonstrated, our proposed 2D-SNet has superiority outstrips closely-related advanced approaches for person re-identification on datasets Market-1501 and CUHK03.