A Preliminary Study of Attribute-aware Semantic Segmentation for Pedestrian Understanding

for Pedestrian Understanding Mahmud Dwi Sulistiyo ,Yasutomo Kawanishi, Daisuke Deguchi, Ichiro Ide, Hiroshi Murase Graduate School of Informatics, Nagoya University, Japan, School of Computing, Telkom University, Indonesia 1.MOTIVATION Semantic segmentation is widely developed in many researches for various purposes. In recent years, several studies have been conducted using various approaches for better classifying a number of distinct objects. Furthermore, semantic information is beneficial for segmentation and segmentation results help more in recognizing semantics. However, only class of each region is considered as the semantic information in traditional semantic segmentation methods. We consider attributes of objects as such additional information. For instance, since the silhouette of a pedestrian is strongly related to the pedestrian's facing orientation, it can help orientation recognition of the pedestrian. We are aiming to collaborate between semantic segmentation and attribute recognition for better results in scene understanding. As the first step, we utilize semantic segmentation results for pedestrian orientation recognition.