sparsity boundaries

Study on Pedestrian Attribute Recognition Identifies Sparsity Wall and Optimizes Edge Deployment

A new study on pedestrian attribute recognition (PAR) addresses extreme class imbalance in large-scale datasets. Researchers identified the "majority negative class cheating trap" and proposed a calibrated Multi-Label Focal Loss configuration. They also defined the "Sparsity Wall," a boundary where global loss reweighting fails, requiring instance-level intervention.