ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Journal Article 포인트 클라우드에서 복셀화 기법에 따른 3D CNN 학습 모델의 XAI 성능 비교 분석
Cited - time in scopus Share share facebook twitter linkedin kakaostory
Authors
김재홍, 서재석, 박기욱, 김광수, 이기훈
Issue Date
2026-09
Citation
소프트웨어포렌식 논문지, v.22, no.3, pp.205-215
ISSN
3092-541X
Publisher
한국소프트웨어감정평가학회
Language
Korean
Type
Journal Article
DOI
https://dx.doi.org/10.29056/jsf.2026.09.18
Abstract
This paper analyzes input channel contributions and spatial activations of a dense 3D CNN for urban point cloud semantic segmentation from an Explainable AI (XAI) perspective. A cropped region of the STPLS3D dataset was voxelized using the Full-Z+RGB-16 representation, consisting of occupancy, normalized relative z-height, R16, G16, and B16, and GradientSHAP, Grad-CAM, and ROAR were applied. The results showed that occupancy had the highest mean absolute SHAP attribution for both Building and Car, while the relative importance of height and color channels differed by class. Grad- CAM activations were observed around occupied voxels as well as near patch boundaries, indicating a possible boundary effect. In the ROAR experiment, at a 40% channel removal rate, the Car IoU after SHAP ranked removal was 20.79%, approximately 15.45 percentage points lower than the 36.24% obtained with random removal.
Keyword
3D CNN, Voxelization, Explainable AI (XAI), Semantic Segmentation
KSP Keywords
3D CNN, Percentage points, Point clouds, Relative importance, Semantic segmentation, Z-height, color channel, removal rate