SAI3D: Zero-Shot 3D Instance Segmentation Approach with SAM
Core Concepts
SAI3D introduces a novel zero-shot 3D instance segmentation approach that leverages geometric priors and semantic cues from the Segment Anything Model (SAM).
Abstract
The paper introduces SAI3D, a zero-shot 3D instance segmentation method.
It combines geometric priors and semantic-aware image segmentation.
SAI3D dissects 3D scenes into geometric primitives for accurate segmentations.
A region-growing algorithm with dynamic thresholding enhances robustness in fine-grained scene parsing.
Empirical evaluations on various datasets demonstrate the superiority of SAI3D over existing methods.
SAI3D
Stats
Empirical evaluations on ScanNet, Matterport3D, and ScanNet++ datasets demonstrate the superiority of our approach.
Quotes
"Our method partitions a 3D scene into geometric primitives, which are then progressively merged into 3D instance segmentations."
"In this work, we investigate how to better leverage geometric priors and multi-view consistency for fine-grained instance segmentation of intricate 3D scenes."