Efficient Deep Learning for Key Challenges in Person Re-identification

This work presents specially designed, computationally efficient deep learning techniques for addressing some key challenges in the area of person re-identification (re-ID). In particular, we focus on scale variations and occlusions, two critical issues that frequently make re-ID tasks, and in a broader sense the problem of real-world surveillance, an extremely challenging goal. This chapter begins by exploring advancements in scale-invariant architectures, such as scale-invariant residual networks and a novel batch-adaptive triplet loss function. These innovations aim to improve feature extraction across different scales and enhance the discriminative power of learned embeddings. Additionally, this chapter introduces a lightweight solution for partial and occluded re-ID using a projective dictionary learning framework, optimized through knowledge distillation from high-performing deep networks. This approach effectively handles incomplete visibility by focusing on regions, which are not occluded, and leveraging a unary-binary dictionary learning model for improved matching accuracy. The proposed methods are rigorously evaluated across several benchmark datasets, demonstrating superior performance in handling both scale variations and occlusions while maintaining computational efficiency. The solutions presented in this chapter, when comprehensively tested against state-of-the-art competitors on a number of publicly available datasets, clearly reveal their efficacy. These solutions offer practical advancements for re-ID systems, with applications in large-scale surveillance and security settings.