Open-set 3D Semantic Segmentation via Transductive Adversarial Prototype Framework

3D Point cloud semantic segmentation is an essential yet challenging task. Most existing methods assume that the training and test point clouds share the same set of object classes - an assumption that often fails in real-world scenarios where the goal is to recognize 3D objects belonging to classes not seen during training. To tackle this challenge, we introduce a transductive Adversarial Prototype Framework (T-APF) for open-set 3D semantic segmentation, which simultaneously identifies points from unseen classes and preserves high segmentation accuracy on seen classes. T-APF integrates four core components: (1) a feature extraction module for capturing point-wise features, (2) a prototypical constraint module that derives a representative prototype for each class, (3) a feature adversarial module based on generative adversarial networks (GANs) to synthesize plausible features for unseen classes, and (4) an unseen-class detection module that generates pseudo-labels for test points during inference. These synthesized features guide the model to learn more discriminative representations and robust prototypes, enhancing its ability to distinguish between seen and previously unobserved semantic categories. The proposed framework is flexible enough to incorporate existing 3D closed-set segmentation backbones, enabling straightforward adaptation to the open-set setting. Extensive experiments on three public benchmarks show that the resulting models consistently achieve significantly better performance than current state-of-the-art methods across most evaluation settings.