Philipp Schröppel received his Diploma degree (equivalent of MSc) in Information Systems Engineering at the TU Dresden in 2018. Since January 2019, he is a PhD student in the Computer Vision Group at the University of Freiburg, headed by Prof. Thomas Brox. His main research interest is 3d reconstruction from images with a focus on generalization across different domains and robust application in real-world scenarios.

AG 1, AG 2, AG 3, INET, AG 4, AG 5, D6, SWS, RG1, MMCI   Info

Expert Audience

English

60 Minutes

Saarbrücken

This talk will be on robust multi-view depth estimation across different domains. In this task, multiple views of the same scene are given. One of these views is the so-called keyview and the other views are called source views. The task is to estimate a depth map for the keyview. As the depth information can be derived from the motion parallaxes between the key- and source views, good generalization across domains should be possible.

In our recent publication "A Benchmark and A Baseline for Multi-view Depth Estimation", we introduce a benchmark for this task, that we term the the Robust Multi-view Depth Benchmark. It is built upon a set of public datasets and allows evaluation on data from different domains. We evaluate recent approaches and find imbalanced performances across domains. Further, we consider a third setting where camera poses are available and the objective is to estimate the corresponding depth maps with their correct scale. We show that recent approaches do not generalize across datasets in this setting. This is because their cost volume output runs out of distribution.

In the talk I will present our findings on the Robust MVD Benchmark in more detail. Further, I will present the accompanying robustmvd framework (https://github.com/lmb-freiburg/robustmvd), which allows usage and evaluation of different multi-view depth estimation models with a common interface.

Mona Linn

+49 681 302 70157

--email hidden

passcode not visible

logged in users only