Abstract

This paper discusses the results of the third edition of the Monocular DepthEstimation Challenge (MDEC). The challenge focuses on zero-shot generalizationto the challenging SYNS-Patches dataset, featuring complex scenes in naturaland indoor settings. As with the previous edition, methods can use any form ofsupervision, i.e. supervised or self-supervised. The challenge received a totalof 19 submissions outperforming the baseline on the test set: 10 among themsubmitted a report describing their approach, highlighting a diffused use offoundational models such as Depth Anything at the core of their method. Thechallenge winners drastically improved 3D F-Score performance, from 17.51% to23.72%.

 

Quick Read (beta)