Distilled Neural Network for Clothes Type Identification

Diverse applications such as criminal forensics and retail customer market segmentation would benefit from the ability to track individuals across multiple cameras. Clothing information is one of the keys to this capability when other identifying information such as the face cannot be reliably extracted. To quickly search for a criminal suspect or re-identify individuals across time and space, filtering by clothing over a given set of cameras for a given time range can dramatically narrow the search. To achieve this capability, we propose a training pipeline that adapts a knowledge distillation neural network to develop a lightweight but effective deep learning model for classifying clothing. The proposed training pipeline produces a small model that achieves 78.70% clothes type classification accuracy on the DeepFashion2 benchmark dataset. An experimental inference time of 0.019 per image on a Nvidia RTX 2070 Super GPU demonstrates that the system has potential for deployment over large-scale camera sensor networks. We expect the new approach to accelerate criminal investigations and enable other analytics requiring reidentification of individuals across multiple cameras at scale.