An Image-based Measuring Technique for the Prediction of Human Body Size

To achieve high prediction accuracy of human body keeps an open issue for decades of years, especially when COVID comes and online retail becomes the major consumption channels. The body measurement is the key to solve cloth matching and recommendation in clothing e-commerce. This paper proposes a practical framework of image-based body measurement, by only taking the user's front and side photos. This framework does not require pure background or precise standing position, and supports manual modification of the measurement results. The framework takes people's height, weight and gender as params to initialize a common body size set, and corrects each part of the set by analyzing the body proportion via the front and side images. The prediction accuracy was tested with the 50 digital models and 10 real people. Results showed that the circumference sizes such as chest, waist, hips, have errors less then 5%, while the length sizes such as arm, leg approach to actual length on net body models. For real people, the errors depend on the wearing clothes. In addition to high accuracy, the method has a rapid process speed, reaching 19QPS on a NVIDIA RTX5000 GPU server.