MVF-PointCLIP: training-free multi-view fusion PointCLIP for zero-shot 3D classification

The remarkable success of Contrastive Language-Image Pretraining (CLIP) in zero-shot 2D vision classification inspires researchers to explore its potential application to zero-shot 3D classification. Some researchers project 3D point clouds into 2D images from multiple views to leverage CLIP. Howeve...

ver descrição completa

Detalhes bibliográficos
Autores: Dai, Jiuqian, Ji, Zhenyan, Xiong, Zechang, Zhu, Guiping, Liu, Hui, Yin, Shen, Armendáriz Íñigo, José Enrique
Tipo de documento: artigo
Estado:Versión aceptada para publicación
Data de publicação:2025
País:España
Recursos:Universidad Pública de Navarra
Repositório:Academica-e. Repositorio Institucional de la Universidad Pública de Navarra
OAI Identifier:oai:academica-e.unavarra.es:2454/55920
Acesso em linha:https://hdl.handle.net/2454/55920
Access Level:Acesso embargado
Palavra-chave:Point cloud
CLIP
Multi-view fusion
Zero-shot
Training-free
Descrição
Resumo:The remarkable success of Contrastive Language-Image Pretraining (CLIP) in zero-shot 2D vision classification inspires researchers to explore its potential application to zero-shot 3D classification. Some researchers project 3D point clouds into 2D images from multiple views to leverage CLIP. However, by doing experiments, we find that this method suffers from two critical drawbacks: (1) noise views existing in multiview depth maps, which provide limited information that may mislead classification; (2) covariance inconsistencies between sample views, which can lead to misclassification when using cosine similarity. To address these issues, we propose a training-free MultiView Fusion PointCLIP (MVF-PointCLIP). It contains a Spatial and Frequency Attention (SFA) module and a Mahalanobis Distance module designed by us. The SFA module automatically assigns importance weights to views, effectively filtering out noisy information. The Mahalanobis Distance module models the distribution of views to tackle covariance inconsistencies. Experimental results verify the superiority of MVF-PointCLIP to SOTA models in zero-shot classification across ModelNet10, ModelNet40, and ScanObjectNN.