Article Dans Une Revue Pattern Recognition and Computer Vision Année : 2025

Target-Specific Domain Adaptation via Geometry-Correlation Prediction for Point Cloud

Résumé

Point cloud datasets suffer from a large domain discrepancy due to variations in data acquisition procedures, sensor perspectives, and realistic noise. To address this issue, domain adaptation technologies have emerged to improve scalability and generalization for point cloud models. We propose a novel method named GeoCo-TSDA (Geometry-Correlation Prediction and Target-Specific Domain Adaptation) which boosts the performance of domain adaptation with a geometry-correlation prediction task and a target-specific self-training strategy. We design a self-supervised task that predicts the geometry correlation, which is obtained by the covariance of the local point clusters and is related to a variety of geometric properties, enabling the model to learn more complete and robust features. Moreover, existing domain adaptation methods commonly focus on aligning feature space among domains. However, owing to the internal distribution gap among domains, merely aligning features at the domain level falls short of achieving optimal performance for the target domain. We tackle this problem by adding a target domain specific training procedure that focuses on further adapting the model to fit the internal distribution of the target domain. For the rationality of our method, we provide theoretical and empirical analysis. For the effectiveness of our method, we conduct experiments on commonly used benchmark PointDA-10 and GraspNetPC-10, and on both datasets our model achieves SOTA performance among point cloud domain adaptation methods and considerable elevation compared to the baseline model.

Fichier non déposé

Dates et versions

hal-05406240 , version 1 (09-12-2025)

Identifiants

Citer

Junqiao Li, Leyan Zhu, Tian Wang, Yuan Xie, Jingang Shi, et al.. Target-Specific Domain Adaptation via Geometry-Correlation Prediction for Point Cloud. Pattern Recognition and Computer Vision, 2025, Lecture Notes in Computer Science, 15034, pp.47-61. ⟨10.1007/978-981-97-8505-6_4⟩. ⟨hal-05406240⟩

Collections

37 Consultations
0 Téléchargements

Altmetric

Partager

  • More