An Empirical Analysis of Optimization Dynamics and Sparsity Boundaries in Large-Scale Pedestrian Attribute Recognition

The paper presents an empirical analysis of optimization dynamics in Pedestrian Attribute Recognition (PAR) using a ResNet-18 backbone, addressing extreme class imbalance in a 109,000-image dataset. By employing Multi-Label Focal Loss with specific hyperparameters (alpha=0.50, gamma=2.0), the authors achieve a Macro F1-score of 62.32%, effectively mitigating the majority negative class cheating trap without additional computational costs. The study introduces the concept of the Sparsity Wall, highlighting a threshold where traditional loss reweighting fails, necessitating instance-level adjustments for effective learning in scenarios with extremely low positive sample fractions.