Deep neural networks for human pose estimation from a very low resolution depth image

The work presented in the paper is dedicated to determining and evaluating the most efficient neural network architecture applied as a multiple regression network localizing human body joints in 3D space based on a single low resolution depth image. The main challenge was to deal with a noisy and coarse representation of the human body, as observed by a depth sensor from a large distance, and to achieve high localization precision. The regression network was expected to reason about relations of body parts based on depth image, and to extract locations of joints, and provide coordinates defining the body pose. The method involved creation of a dataset with 200,000 realistic depth images of a 3D body model, then training and testing numerous architectures including feedforward multilayer perceptron network and deep convolutional neural networks. The results of training and evaluation are included and discussed. The most accurate DNN network was further trained and evaluated on an augmented depth images dataset. The achieved accuracy was similar to a reference Kinect algorithm results, with a great benefit of fast processing speed and significantly lower requirements on sensor resolution, as it used 100 times less pixels than Kinect depth sensor. The method was robust against sensor noise, allowing imprecision of depth measurements. Finally, our results were compared with VGG, MobileNet, and ResNet architectures.