Topology-Aware Self-Training Framework Improves Semi-Supervised 3D Shape Segmentation

A topology-aware multi-task self-training framework improves semi-supervised 3D shape segmentation by controlling confirmation bias from noisy pseudo-labels, according to research by Zhichao Zhang and Miao Zhu published in Discover Applied Sciences. In the domain of 3D shape segmentation, computational models assign semantic labels to individual mesh facets across geometric objects. However, standard fully supervised segmentation methods require extensive, costly fine-grained annotations that are difficult to produce at scale. While semi-supervised self-training offers a viable path to reduce this manual annotation burden, confirmation bias stemming from noisy pseudo-labels frequently destabilizes network training and severely degrades part boundaries between segmented regions.

To address these training instabilities, the researchers developed a topology-aware multi-task self-training framework built upon a teacher-student network architecture. Within this system, an exponential-moving-average teacher network continuously generates pseudo-label targets to guide the optimization of a student network. Because raw pseudo-labels introduce substantial confirmation bias when erroneous predictions reinforce themselves, the framework integrates three complementary components designed specifically to manage and control pseudo-label noise throughout the training process: a topology consistency module, an auxiliary decoder, and an iterative graph-guided post-processing mechanism.

Each of the three architectural components operates through a distinct mechanism to suppress pseudo-label noise across the network pipeline. The topology consistency module aligns predictions made by the student network with a neighborhood-smoothed teacher distribution, ensuring consistent label predictions across spatially connected facets. Between successive training iterations, the graph-guided post-processing routine refines ambiguous part boundaries where facet classifications are uncertain. Meanwhile, the auxiliary decoder utilizes the complete set of pseudo-labels without directly supervising the primary main decoder. Because the main and auxiliary decoders share a common encoder, pseudo-label supervision actively influences and enriches the shared feature representation without directly adjusting the parameters of the main decoder.

Ablation results reported on the COSEG benchmark demonstrated the specific quantitative contributions of each individual component within the framework. In these ablation experiments, incorporating the topology consistency module produced an improvement of 4.0 percentage points over the baseline setup. Integrating the auxiliary learning mechanism contributed a performance gain of 2.3 percentage points, while applying the graph-guided post-processing stage delivered an additional increase of 2.2 percentage points. These isolated improvements confirm that neighborhood smoothing, decoupled auxiliary decoding, and boundary refinement each provide measurable advantages in combating pseudo-label degradation.

Beyond the ablation studies on COSEG, broader experimental evaluations across the PSB, COSEG, ShapeNetCore, and ShapeNetPart benchmark datasets demonstrated competitive segmentation performance under limited supervision. The researchers further evaluated the framework through targeted noise-corruption experiments as well as tests examining varying labeled-data proportions. These additional investigations demonstrated that combining the topology consistency and auxiliary decoder modules effectively reduces performance degradation when pseudo-labels are imperfect, establishing that the architecture remains robust even when supervision is constrained and pseudo-label targets contain noise.

Subject of Research: Semi-supervised 3D shape segmentation using topology-aware multi-task self-training

Article Title: Topology aware multi task self training for semi supervised 3D shape segmentation

Article References: Zhang, Z., & Zhu, M. (2026). Topology aware multi task self training for semi supervised 3D shape segmentation. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09766-9

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09766-9

Keywords: 3D shape segmentation, semi-supervised learning, self-training, mesh facets, pseudo-labels, confirmation bias, teacher-student architecture, exponential moving average, topology consistency, auxiliary decoder, graph-guided post-processing, shared encoder, part boundaries, COSEG dataset, PSB dataset, ShapeNetCore, ShapeNetPart, noise corruption