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

3D shape segmentation assigns semantic labels to mesh facets, but fully supervised methods require costly fine-grained annotations. Semi-supervised self-training can reduce this burden, although confirmation bias from noisy pseudo-labels often destabilizes training and degrades part boundaries. We propose a topology aware multi task self training framework built on a teacher-student architecture. An exponential-moving-average teacher generates targets for a student network, while three complementary components control pseudo-label noise: a topology consistency module aligns student predictions with a neighborhood-smoothed teacher distribution, an auxiliary decoder uses all pseudo-labels without directly supervising the main decoder, and graph-guided post-processing refines ambiguous boundaries between iterations. The main and auxiliary decoders share an encoder, so pseudo-label supervision influences the shared representation but not the parameters of the main decoder directly. Ablation results on COSEG show improvements of 4.0, 2.3, and 2.2 percentage points from topology consistency, auxiliary learning, and post-processing, respectively. Experiments on PSB, COSEG, ShapeNetCore, and ShapeNetPart show competitive performance under limited supervision. Additional noise-corruption and labeled-data-proportion experiments demonstrate that the combined topology and auxiliary modules reduce performance degradation when pseudo-labels are imperfect.