Fully convolutional architectures for multi-part body segmentation

Abstract

[en] Since the appearance of the baseline Fully Convolutinal Network (FCN), convolution architectures usage has spread widely among Deep Neural Networks: from classification tasks to object tracking, they are found ubiquitously in the Deep Learning field. In this study, three different convolutional architectures are studied with regard its application to the semantic segmentation of the human body: ICNet, a different resolution cascade network, SegNet, a encoder-decoder network, and Stacked Hourglass, a specially purposed network for the human body. For this purpose, the SURREAL (Synthetic hUmans foR REAL tasks) dataset, which consists of synthetically rendered but realistic images of people, is used. As a result, is shown that the best performing network for this task is the Stacked Hourglass. Due to its continuous refinement of the output and the use of the full network for inference a 55.3% mIoU is achieved on the 24 body part dataset.