Manuscript received May 7, 2026; revised May 15, 2026; accepted June 24, 2026; published November 14, 2026.

Abstract—Face segmentation is crucial for various facial analysis tasks. Although significant advancements have been made in deep learning-based 2D face segmentation, detailed part segmentation of 3D faces from point clouds remains largely unexplored due to annotation difficulties and high computational costs. In this paper, we propose an efficient approach for 3D facial part segmentation. Employing an efficient transformer-based model, we introduce two parameter-free lightweight modifications to improve accuracy without sacrificing computational speed: SimAM (Simple Attention Module), an attention module for salient feature enhancement, and BACE (Boundary Aware Cross Entropy), a boundary-aware loss function utilizing frozen sparse convolution with no learnable parameters. We evaluate our method on a private facial dataset that includes complex geometries, such as cleft lip deformities and various facial expressions. The experimental results demonstrate that SimAM and BACE work together to handle these complexities, leading to a statistically significant improvement. Furthermore, we validate the generalizability of the proposed modules using a dataset of 3D objects. Our method yields performance improvements for generic 3D objects when evaluated via per-sample mIoU to prevent unfair penalties from absent classes. In conclusion, integrating SimAM and BACE provides an effective and practical approach to boosting 3D part segmentation performance with minimal computational overhead.

Keywords—3D point cloud, part segmentation, lightweight modules, boundary aware loss, cleft lip

Cite: Narumi Kihara, Namiko Kimura-Nomoto, Takako Okawachi, Guangxu Li, Norifumi Nakamura, and Tohru Kamiya, "Boosting 3D Part Segmentation Performance with Parameter-free Lightweight Modules," Journal of Image and Graphics, Vol. 14, No. 6, pp. 987-996, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).