Geometric-Aware Hypergraph Reasoning for Novel Class Discovery in Point Cloud Segmentation

📄 CVPR 2026 ( Accepted)

[Paper]

Overview

Novel Class Discovery (NCD) in point cloud segmentation aims to automatically identify and segment unlabeled novel categories by transferring knowledge from known classes. Existing approaches mainly rely on pairwise associations, which limits their capability to model complex inter-class relationships.

We propose GAHR (Geometric-Aware Hypergraph Reasoning), a novel framework that introduces high-order relational reasoning into 3D Novel Class Discovery. Instead of modeling binary relations, GAHR constructs a prototype-based hypergraph that enables collaborative reasoning among multiple known classes for discovering unseen categories.

As illustrated in Figure 2, point cloud features are first transformed into Geometric-Aware Prototypes, which serve as hypergraph nodes. Hyperedges dynamically connect related prototypes according to geometric and semantic similarities, allowing high-order knowledge propagation for robust novel class inference. CVPR2026 (41)

---

Figure 2. Overall architecture of GAHR. Point cloud features are clustered into geometric-aware prototypes, which form nodes in a dynamically updated hypergraph for collaborative novel class reasoning.

Installation

Please refer to the environment setup of NOPS for dependency installation (e.g., MinkowskiEngine, CUDA, PyTorch, etc.).

Our implementation follows the same backbone and training pipeline, adopting MinkowskiUNet-34C with theAdamW optimizer and a multi-step learning rate schedule.

Data preparation

Please follow the official instructions from SemanticKITTI and SemanticPOSS to download the data. Afterward, structure the folders as follows (the root path should match the path_to_data_shown_in_yaml_config in your yaml config file):

./
├── configs/
├── scripts/
├── train.py
├── test.py
└── path_to_data_shown_in_yaml_config/
      └── sequences
            ├── 00/
            │   ├── velodyne/
            │   │     ├── 000000.bin
            │   │     ├── 000001.bin
            │   │     └── ...
            │   └── labels/
            │         ├── 000000.label
            │         ├── 000001.label
            │         └── ...
            ├── 01/
            └── ...

Commands

Train

python train.py -s 00 --dataset SemanticPOSS --offline --epoch 10 --use_scheduler --lam 1 --lam_region 1 --gamma 1 --alpha 1 --gamma_decrease 0.5 --smooth_bound 10 --ak_bound 0.005 --dbscan 0.5

python train.py -s 00 --dataset SemanticKITTI --offline --epoch 10 --use_scheduler --lam 1 --lam_region 1 --gamma 1 --alpha 1 --gamma_decrease 0.5 --smooth_bound 10 --ak_bound 0.005 --dbscan 0.5

Citation

If you find our work useful, please cite our paper:

@inproceedings{GAHR2026,
  title={Geometric-Aware Hypergraph Reasoning for Novel Class Discovery in Point Cloud Segmentation},
  author={Anonymous},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}