ColAtt‐Net: In Reducing the Ambiguity of Pedestrian Orientations on Attribute‐Aware Semantic Segmentation Task

Abstract

AbstractSemantic segmentation has become one of the trending topics in the world of computer vision and deep learning. Recently, due to an increasing demand to solve a semantic segmentation task simultaneously with attribute recognition of objects, a new task named attribute‐aware semantic segmentation has been introduced. Since the task requires to handle pixel‐wise object class estimation with its attributes such as a pedestrian's body orientation, previous works had difficulties to handle ambiguous attributes such as body orientations in object‐level, especially when segmenting the pedestrians with their attributes correctly. This paper proposes the ColAtt‐Net that is an attribute‐aware semantic segmentation model augmented by a column‐wise mask branch to predict the pedestrians' orientations in the horizontal perspective of the input image. We firmly assume that the pedestrians captured by a car‐mounted camera are distributed horizontally so that for each column of the input image, the pedestrian pixels can be labeled with one orientation uniformly. In the proposed method, we split the output of the base semantic segmentation model into two branches; one branch for segmenting the object categories, while the other one, as the novel column‐wise attribute branch, is to map the recognition of pedestrian's orientations that are distributed horizontally. This method successfully enhances the performance of attribute‐aware semantic segmentation by reducing the ambiguity on segmenting the pedestrian's orientation. Improvements on the pedestrian orientation segmentation are confidently shown by the proposed method in the experimental results, both in quantitative and qualitative views. This paper also discusses how the improved performance becomes an advantage in the autonomous driving system. © 2020 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.