Article citationsMore>>

Wu, X., Jiang, L., Wang, P., Liu, Z., Liu, X., Qiao, Y., et al. (2024) Point Transformer V3: Simpler, Faster, Stronger. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, 16-22 June 2024, 4840-4851.
https://doi.org/10.1109/cvpr52733.2024.00463

has been cited by the following article:

• TITLE: Geo-Refined Point Transformer: Coordinate-Aware Excitation and Positional Upsampling for 3D Scene Segmentation

  AUTHORS: Jingwei Lu, Yi Zhang

  KEYWORDS: Geometric Information Vacuum, Coordinate-Aware Feature Excitation, Position-Aware Upsampling, Efficiency-Accuracy Trade-Off

  JOURNAL NAME: Journal of Computer and Communications, Vol.14 No.1, January 19, 2026

  ABSTRACT: As a work exploring the existing trade-off between accuracy and efficiency in the context of point cloud processing, Point Transformer V3 (PTV3) has made significant advancements in computational efficiency through its innovative point cloud serialization strategy. However, this optimization for computational efficiency comes at the cost of sacrificing high-fidelity perception of fine-grained local geometric structures, thereby introducing a limitation termed “geometric information vacuum” in the model. To address this issue, our work proposes a coordinate-aware feature activation module, which enhances the model’s sensitivity to spatial locations by dynamically calibrating feature channel responses using the 3D absolute coordinates of points during the encoder stage. Furthermore, our work designs a position-aware upsampling mechanism that accurately restores the geometric details smoothed out during downsampling by learning a feature compensation term associated with the relative positions of points within voxels during the decoder stage. Experiments on 3D point cloud segmentation on S3DIS and ScanNe t v2 show that the Geo-PT model proposed in this study achieves better performance than PTv3 with negligible additional computational cost.