360 Layout Estimation via Orthogonal Planes Disentanglement and Multi-View Geometric Consistency Perception

Area of Science:

• Computer Vision
• Machine Learning
• Computational Geometry

Background:

• Existing panoramic layout estimation methods suffer from imprecise room boundary recovery due to vertical compression.
• Data-driven approaches require laborious and time-consuming manual annotations, limiting scalability.

Purpose of the Study:

• To develop a more accurate and data-efficient method for panoramic layout estimation.
• To address the semantic ambiguity and annotation burden in current solutions.

Main Methods:

• Proposing an orthogonal plane disentanglement network (DOPNet) to create distortion-free, semantics-clean representations.
• Introducing an unsupervised adaptation technique for horizon-depth and ratio representations.
• Utilizing an optimization strategy for decision-level layout analysis and 1D cost volume construction for multi-view feature aggregation.

Main Results:

• DOPNet effectively disentangles ambiguous semantics, leading to improved layout recovery.
• The unsupervised adaptation technique leverages geometric consistency across multiple perspectives, generating reliable pseudo-labels.
• The proposed solution demonstrates superior performance over state-of-the-art models in both monocular and multi-view layout estimation tasks.

Conclusions:

• The developed method significantly enhances the accuracy and efficiency of panoramic layout estimation.
• The orthogonal plane disentanglement and unsupervised adaptation techniques offer a promising direction for future research in scene understanding.