Abstract
Pedestrian Attribute Recognition (PAR) focuses on identifying variousattributes in pedestrian images, with key applications in person retrieval,suspect re-identification, and soft biometrics. However, Deep Neural Networks(DNNs) for PAR often suffer from over-parameterization and high computationalcomplexity, making them unsuitable for resource-constrained devices.Traditional tensor-based compression methods typically factorize layers withoutadequately preserving the gradient direction during compression, leading toinefficient compression and a significant accuracy loss. In this work, wepropose a novel approach for determining the optimal ranks of low-rank layers,ensuring that the gradient direction of the compressed model closely alignswith that of the original model. This means that the compressed modeleffectively preserves the update direction of the full model, enabling moreefficient compression for PAR tasks. The proposed procedure optimizes thecompression ranks for each layer within the ALM model, followed by compressionusing CPD-EPC or truncated SVD. This results in a reduction in model complexitywhile maintaining high performance.