Scale-equivariant convolution for semantic segmentation of depth image

Accurate understanding of the environment is crucial for autonomous driving and robot automation. Depth sensors, including light detection and ranging and depth cameras, are attracting attention. It is practical to treat the depth information in a depth image form. With the progress in Artificial Intelligence, many deep neural networks have been proposed for the segmentation of depth images. However, no method has focused on the difference in scale within an image caused by a 3-dimensional to 2-dimensional projection. We proposed a new scale-equivariant convolution method that focuses on the relationship between the object distance and scale ratio in the image.