An UNet-Based Head Shoulder Segmentation Network

Within the rapidly developing field of computer vision, pedestrian detection is a fundamental and challenging task for both industry and academia. However, object segmentation information can help the network to capture the attention of the model during training. In this paper, we propose a head-shoulder segmentation network based on modified U-Net network. The architecture consists of a contracting path to capture information from a lower layer and a symmetric expanding path to enable precise localization. The proposed model aims to effectively segment the head-shoulder portion of pedestrian without a huge annotated training sample. Segmentation of a random image takes less than a second on NVIDIA GTX 1070. This paper will show the mean IOU and some segmentation results to prove effectiveness of this model.