Garg Divyansh ; Wang Yan ; Hariharan Bharath ; Campbell Mark ; Weinberger Kilian Q. ; Chao Wei-Lun

Abstract

Existing approaches to depth or disparity estimation output a distributionover a set of pre-defined discrete values. This leads to inaccurate resultswhen the true depth or disparity does not match any of these values. The factthat this distribution is usually learned indirectly through a regression losscauses further problems in ambiguous regions around object boundaries. Weaddress these issues using a new neural network architecture that is capable ofoutputting arbitrary depth values, and a new loss function that is derived fromthe Wasserstein distance between the true and the predicted distributions. Wevalidate our approach on a variety of tasks, including stereo disparity anddepth estimation, and the downstream 3D object detection. Our approachdrastically reduces the error in ambiguous regions, especially around objectboundaries that greatly affect the localization of objects in 3D, achieving thestate-of-the-art in 3D object detection for autonomous driving. Our code willbe available at https://github.com/Div99/W-Stereo-Disp.