CNN Architectures for Human Pose Estimation from a Very Low Resolution Depth Image

The paper is dedicated to proposing and evaluating a number of convolutional neural network architectures for calculating a multiple regression on 3D coordinates of human body joints tracked in a single low resolution depth image. The main challenge was to obtain a high precision in case of a noisy and coarse scan of the body, as observed by a depth sensor from a large distance. The regression network was expected to reason about relations of body parts based on depth image, and to extract locations of joints. The method involved creation of a dataset with 200,000 realistic depth images of a 3D body model, then training and testing numerous CNN architectures. The results are included and discussed. The achieved accuracy was similar to a reference Kinect algorithm results, with a great advantage of fast processing speed and significantly lower requirements on sensor resolution, as it used 100 times less pixels than Kinect depth sensor.