SPREAD-GS: Scale-Progressive Representation Extraction and Detailing for 3D Gaussian Splatting

Yifan Liu, Zheng Xiao, Lingfan Zheng, Yue Liu, Jianbin Jiao, Yanzhao Zhou

Keywords: 3D Gaussian Splatting, multi-scale representation, feature propagation, 3D object classification

TL;DR: SPREAD-GS leverages scale-progressive sampling and hierarchical detail propagation to improve texture- and geometry-aware 3D object classification.

Abstract:

3D Gaussian Splatting (3DGS) has recently shown remarkable capability for high-fidelity scene reconstruction. However, its potential for object recognition remains under-explored. Existing approaches often extract 2D features from multi-view images and embed them into 3DGS, which limits the joint use of 3DGS geometric and appearance information.
To fully exploit both the structural and fine-grained details encoded in 3DGS, we propose Scale-Progressive Representation Extraction And Detailing for 3D Gaussian Splatting (SPREAD-GS), a framework for object classification that combines scale-aware sampling with detail-preserving feature propagation.
SPREAD-GS has two key modules: Scale-Progressive Sampling (SPS), generating multi-scale subsets by progressively narrowing the visible region, and SpreadNet, encoding these subsets and propagating details across scales through noise-augmented feature upsampling. On the texture-rich MACGS dataset, SPREAD-GS achieves 93.93\% overall accuracy, improving the SOTA by 2.02\%. On the geometry-centric ModelNet40GS, it matches the SOTA while significantly reducing parameters.
These results demonstrate the effectiveness of scale-progressive sampling and detail-preserving feature propagation for 3DGS recognition.
These results demonstrate the effectiveness of scale-progressive sampling and detail-preserving feature propagation for 3DGS recognition. Our code is available at https://anonymous.4open.science/r/noname-64BE.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 2493