An Empirical Analysis of Optimization Dynamics and Sparsity Boundaries in Large-Scale Pedestrian Attribute Recognition
Researchers have identified a significant challenge in pedestrian attribute recognition (PAR) caused by extreme class imbalance in large datasets like PETA and PA-100K. This imbalance leads to a phenomenon termed the 'majority negative class cheating trap,' where standard optimization methods suppress rare attributes. By systematically analyzing Multi-Label Focal Loss hyperparameters on a ResNet-18 backbone, a calibrated configuration was found to match baseline performance while improving rare attribute detection and convergence. The study also defines the 'Sparsity Wall,' a boundary below which global loss reweighting becomes ineffective, necessitating instance-level interventions. AI
IMPACT Identifies a critical data imbalance issue in AI recognition tasks and proposes a loss function tuning method to mitigate it.