Evaluation of MC-CNN Based Stereo Matching Pipeline for the CO3D Earth Observation Program

Véronique Defonte, Loïc Dumas, Myriam Cournet, Emmanuelle Sarrazin

Abstract: When it comes to 3D reconstruction, stereo matching step stands in the way of the light. But just as essential as it is, it remains very critical. Even more so for the CO3D world wide mission. 3D will be rendered for every continents, increasing the importance of generalizing well over different landscapes. This paper then focuses on the benefits of using MC-CNN neural network as stereo matching costs, instead of the Census adopted so far on CO3D studies. We compare them both over different areas. We train MC-CNN on Middlebury dataset and a stereo satellite dataset. We show that MC-CNN matching costs are much more trustworthy than Census ones and introduce a measure of ambiguity to illustrate this behavior. To help reproduce our results, we publicly release our work as a contribution to Pandora.