💻 computer science

Category-aware and Centroid-Guided Point Transformer for 3D Ear Region Segmentation

This paper proposes a category-aware and centroid-guided point transformer built upon Point Transformer v3 to address class imbalance and boundary blurring in 3D ear region segmentation, achieving superior accuracy in both large and tiny subregions through a class-balanced attention module, dynamic loss function, and centroid-guided post-processing.

Original authors: Li Yuan, Wenhao Zuo, Jiangyun Li, Wenjiang Huang

Published 2026-09-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Li Yuan, Wenhao Zuo, Jiangyun Li, Wenjiang Huang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The human ear is a complex landscape of curves, ridges, and deep hollows, a structure so intricate that it has long challenged engineers trying to design comfortable headphones, hearing aids, or medical devices that fit perfectly. To build something that sits well on the ear, designers need to know the exact shape and size of every tiny part, from the outer rim to the deepest inner chambers. For decades, researchers have tried to map these shapes using flat photographs, but a two-dimensional picture cannot capture the true depth and curvature of a three-dimensional object. It is like trying to understand the shape of a mountain range by looking at a single shadow; the depth is lost, and the details blur. To solve this, scientists have turned to 3D point clouds, which are digital collections of millions of tiny dots that map the surface of an object in space, preserving every curve and angle. However, even with these detailed maps, teaching computers to recognize the difference between a large outer section of the ear and a tiny, specific spot inside it has proven difficult. The computer tends to ignore the small, rare details because they are overwhelmed by the sheer volume of data from the larger areas, much like a spotlight that is too bright to see the small flowers growing in the shadow of a giant tree.

A team of researchers at the University of Science and Technology Beijing and the Chinese Academy of Sciences has developed a new method to help computers see these tiny details clearly. They created a specialized digital model designed to segment, or separate, the human ear into its distinct regions with high precision. Their approach focuses on a specific problem: the ear contains eight major regions, such as the earlobe and the outer rim, which cover most of the surface, and eight much smaller, critical regions hidden deep inside the main hollow, which correspond to specific points used in traditional medicine. Because the small regions have far fewer data points than the large ones, standard computer models often mistake them for their larger neighbors or fail to identify them entirely. The researchers built their solution on an existing advanced system known as a Point Transformer, which is already skilled at understanding 3D shapes, but they added three new layers of intelligence to fix the imbalance.

First, the team introduced a mechanism that forces the computer to pay extra attention to the small, rare parts of the ear. In the digital model, this works by assigning a special weight to the tiny regions, effectively telling the system, "Do not overlook these." This ensures that the features of the small areas are amplified during the learning process, making them stand out against the background of the larger ear sections. Second, they adjusted the way the computer learns from its mistakes. Instead of treating every error the same, the system now penalizes mistakes involving the small regions much more heavily. If the computer misidentifies a tiny spot inside the ear, the penalty is severe, pushing the model to correct that specific error immediately. This dynamic adjustment ensures the model does not focus solely on the easy, large areas. Finally, the researchers added a finishing step after the computer makes its initial guess. Because the computer sometimes gets a little too excited about the small regions and marks too many dots as belonging to them, the team uses a geometric rule to clean up the result. They calculate the exact center point of the predicted small area and then select only the nearest twenty-five points to that center to be the final, official shape. This ensures the identified region is compact, round, and true to the anatomy, removing any scattered or noisy points that might have been included by mistake.

The researchers tested their new system on a dataset of 500 human ear models, each containing nearly 30,000 points. The results showed a significant improvement over previous methods. While older models struggled to distinguish the tiny inner regions, often blurring their boundaries or missing them completely, the new system achieved a high level of accuracy, correctly identifying the average overlap between the predicted shape and the true shape at 0.8387. More importantly, it correctly classified the average point across all regions with an accuracy of 0.9510. When compared to other leading technologies, the new method produced much sharper boundaries and far fewer errors, particularly in the crowded, complex inner parts of the ear. Visual comparisons showed that while other models left the small regions fragmented or confused with the surrounding tissue, this new approach produced clean, continuous shapes that matched the real anatomy perfectly.

The significance of this work lies in its ability to make digital ear models reliable enough for real-world design. By solving the problem of class imbalance—where large areas drown out small ones—the researchers have provided a tool that can accurately measure the geometric size of every part of the ear. This precision is essential for creating ear-worn devices that are not only functional but also comfortable for long-term use. The team has made their code and data publicly available, allowing other scientists to verify their findings and build upon their work. In the future, this technology could be integrated with medical diagnostics to offer even more detailed analysis of the ear, supporting the design of personalized devices and improving the accuracy of auricular analysis for healthcare applications. The study demonstrates that with the right combination of attention, learning rules, and geometric refinement, computers can finally learn to see the smallest details of the human body with the same clarity as the largest.

Technical Summary: Category-aware and Centroid-Guided Point Transformer for 3D Ear Region Segmentation

1. Problem Statement

The accurate segmentation of 3D ear point clouds is critical for the design of comfortable and functional ear-worn devices (e.g., headphones, hearing aids, ear massagers) and for auricular analysis. The human ear consists of eight major regions (e.g., helix, concha, earlobe) and numerous smaller sub-regions (e.g., specific acupoint areas within the concha).

Existing point cloud segmentation methods face significant challenges in this domain due to:

Severe Class Imbalance: Large regions (like the helix) contain the majority of points, while tiny sub-regions (like specific acupoints) have very few points. This leads models to bias toward large regions, causing misclassification and "erosion" of small regions.
Geometric Complexity: Small regions are closely adjacent to larger regions with complex curved surfaces and local concave-convex features.
Boundary Fragmentation: Standard methods often produce fragmented boundaries and irregular shapes for small regions, failing to capture the precise anatomical structures required for device fitting and medical analysis.
Limitations of 2D Approaches: While 2D image-based methods exist, they lose depth information and struggle with the complex 3D topology of the ear, particularly for fine-grained localization.

2. Methodology

The authors propose a Category-aware and Centroid-Guided Point Transformer framework built upon the Point Transformer v3 baseline. The architecture consists of four main components: preprocessing, an encoder, a decoder, and a segmentation head, augmented by three specific innovations to address the class imbalance and geometric precision issues.

A. Class-Balanced Attention Module

Located after the decoder, this module aims to enhance the feature representation of small region categories (labels 8–15) which are often underrepresented.

Mechanism: It utilizes a learnable category weight vector. Large regions (0–7) are assigned a dynamic weight $\alpha$ , while small regions (8–15) are assigned a dynamic weight $\beta$ .
Dynamic Adjustment: Through L2 and sparsity regularization, the model learns to increase $\beta$ relative to $\alpha$ as the network depth increases. This shifts the model's attention from global geometric structures in shallow layers to high-level semantic features of small regions in deeper layers.
Feature Enhancement: The module multiplies decoder features by an attention factor derived from these weights, amplifying the features of small regions to prevent them from being "eroded" by adjacent large regions.

B. Dynamic Category-Aware Loss Function

To further guide the training process toward small regions, a composite loss function is employed:
$L_{total} = \lambda_{CE}L_{CE} + \lambda_{Lovasz}L_{Lovasz} + \lambda_{DCA}L_{DCA} + \lambda_{CBR}L_{CBR}$

Dynamic Category-Aware Loss ( $L_{DCA}$ ): This loss combines partition weights (assigning a higher weight $\theta$ to small regions) with dynamic weights based on the inverse frequency of points in the current batch. This ensures that misclassified small regions incur a significantly higher penalty, directing optimization toward these difficult categories.
Category Balance Regularized Loss ( $L_{CBR}$ ): An auxiliary loss that constrains the learnable weights in the attention module, preventing them from becoming excessively large while ensuring small regions receive sufficient attention.

C. Centroid-Guided Post-Processing

Even with improved feature learning, direct predictions for small regions can be spatially dispersed or fragmented. The authors propose a post-processing strategy to refine these results:

Centroid Calculation: For each predicted small region category, the geometric centroid ( $\mu_c$ ) of the predicted point set is calculated.
Nearest Neighbor Selection: Using K-Nearest Neighbors (KNN), the 25 points closest to the calculated centroid are selected and assigned to that specific small region category. This strategy leverages the anatomical prior that small ear regions are spatially concentrated around a center, ensuring compact and geometrically regular segmentation results by re-selecting the most representative point set within the neighborhood of the centroid.

3. Key Contributions

Novel Architecture: The integration of a Class-Balanced Attention Module into Point Transformer v3 to dynamically adjust feature weights, specifically targeting the underrepresentation of small ear regions.
Specialized Loss Functions: The introduction of a Dynamic Category-Aware Loss and a Category Balance Regularized Loss to mitigate training bias and enforce focus on minority classes during optimization.
Geometric Post-Processing: A Centroid-Guided Post-Processing method that corrects spatial dispersion in small region predictions, ensuring boundary integrity and compactness without relying on general morphological operations.
Comprehensive Evaluation: Extensive experiments on a dataset of 500 ear point cloud samples (derived from the York Ear Model) demonstrating superior performance in both large-region delineation and fine-grained small-region localization.

4. Experimental Results

The proposed method was evaluated on a dataset of 500 ear point clouds (upsampled to 29,800 points), labeled with 8 major regions and 8 small acupoint regions. The model was compared against PointNet++, Point Transformer V2, and Point Transformer V3.

Quantitative Performance:
The proposed method achieved the following metrics, outperforming all baselines:

mIoU (Mean Intersection over Union): 0.8387 (vs. 0.7375 for Point Transformer V3).
mAcc (Mean Accuracy): 0.9510 (vs. 0.8334 for Point Transformer V3).
allAcc (Overall Accuracy): 0.9744.
mPrecision: 0.8744.

Qualitative Observations:

Ablation Studies: The study confirmed that the Class-Balanced Attention module alone improved mIoU to 0.8213. The addition of the Dynamic Loss and Post-Processing further refined the results, with the full model achieving the peak mIoU.
Visual Results: The proposed method successfully segmented all eight major regions with precise anatomical alignment and achieved complete, non-fragmented segmentation of the eight small acupoint regions. In contrast, baselines exhibited severe missegmentation, boundary blurring, and under-segmentation of small regions.

5. Significance and Claims

The paper claims that the proposed framework offers a valuable solution for 3D ear region segmentation, directly addressing the critical challenge of class imbalance in multi-scale point cloud data.

Application Impact: The improved segmentation accuracy and boundary clarity provide reliable geometric support for the design of ear-worn devices (ensuring comfort and fit) and auricular analysis (supporting medical or biometric applications).
Methodological Value: The work demonstrates that combining feature-level attention mechanisms, dynamic loss weighting, and geometry-aware post-processing creates a complete optimization loop that significantly enhances fine-grained segmentation performance in tasks with severe scale disparities.
Reproducibility: The authors have made the code and experimental data publicly available to support further research and replication.

The authors conclude that this approach effectively bridges the gap between large-scale structural understanding and fine-grained local detail, offering a robust foundation for future work in personalized ear-worn device design and intelligent detection.