The paper addresses the inherent ambiguity issue in existing 360° room layout estimation datasets, where the ground truth annotations can be either of an "enclosed" type that stops at ambiguous regions or an "extended" type that encompasses all visible areas.
To tackle this challenge, the authors propose a novel Bi-Layout model that can simultaneously predict two distinct layout types. The key innovations are:
This unique architectural design allows the model to be compact and efficient while addressing the ambiguity issue. The authors also introduce a new "disambiguate" metric to quantitatively evaluate the model's ability to handle ambiguous annotations without the need for manual correction.
Extensive experiments on the MatterportLayout and ZInD datasets demonstrate that the proposed Bi-Layout model outperforms state-of-the-art methods, especially on subsets with significant ambiguity. The model can also inherently detect ambiguous regions by comparing the two layout predictions.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Yu-Ju Tsai,J... at arxiv.org 04-16-2024
https://arxiv.org/pdf/2404.09993.pdfDeeper Inquiries