Core Concepts
A novel method for decomposing 3D scenes into individual objects and backgrounds with minimal human interaction, by integrating the Segment Anything Model (SAM) with hybrid implicit-explicit neural surface representations and a mesh-based region-growing technique.
Abstract
The paper presents Total-Decom, a novel method for decomposed 3D scene reconstruction from multi-view images with minimal human interaction. The key aspects are:
Implicit Neural Surface Reconstruction:
Employs an implicit neural surface representation to achieve dense and complete 3D reconstruction from images.
Integrates object-aware information by distilling image features from the Segment Anything Model (SAM).
Disentangles foreground and background geometry using geometric priors and regularization.
Interactive Decomposition:
Extracts an explicit mesh surface to provide geometry information for better decomposition and efficient rendering.
Leverages the SAM decoder and rendered SAM features to convert a single user click into a dense object mask, enabling interactive control over decomposition granularity.
Mesh-based Region Growing:
Proposes a new mesh-based region-growing algorithm that leverages feature similarities, geometry topology, and object boundaries derived from SAM to accurately extract object surfaces.
Requires minimal human annotations, typically just one click per object on average.
The method is extensively evaluated on benchmark datasets, demonstrating its ability to decompose complex scenes into individual objects with high accuracy, outperforming state-of-the-art approaches. The decomposed 3D reconstruction also enables various downstream applications such as animation and scene editing.
Stats
The paper does not provide any specific numerical data or statistics. The key results are presented through qualitative comparisons and quantitative evaluations of reconstruction accuracy.
Quotes
"Our approach seamlessly integrates the Segment Anything Model (SAM) with hybrid implicit-explicit neural surface representations and a mesh-based region-growing technique for accurate 3D object decomposition."
"Total-Decom requires minimal human annotations while providing users with real-time control over the granularity and quality of decomposition."