Geometric Inconsistency Localization in Multi-View Image Sets
DeformView and DEFECt3R localize geometric inconsistencies in wide-baseline multi-view images for forensics.
Novel view synthesis can create realistic views that are not geometrically consistent across viewpoints. The authors release DeformView, a wide-baseline multi-view dataset with pixel-level inconsistency annotations, and find that NVS consistency scores localize forensic inconsistencies poorly. Their classifier DEFECt3R uses cross-view features and hard negatives from consistent but deformed views to improve localization and reduce false positives. Code and the dataset are publicly available.
- DeformView adds pixel-level geometric inconsistency labels for wide-baseline pairs.
- NVS consistency scores transfer poorly to forensic localization.
- DEFECt3R uses cross-view features and hard negatives to cut false positives.
- Feature representations and correspondence quality both affect localization.
From the original article
1 min read at arxiv.orgNovel view synthesis (NVS) models can produce realistic new views of the same scene from different viewpoints. However, these generated views are not always geometrically consistent with one another. Multi-view (MV) consistency has shown promise as a tool for evaluating these NVS models.
By Xander Staelens · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.