ProOOD: Prototype-Guided Out-of-Distribution 3D Occupancy Prediction

We present ProOOD, a lightweight plug-and-play method that addresses long-tailed class bias and out-of-distribution (OOD) detection in 3D semantic occupancy prediction, via three core designs: (1) Prototype-Guided Semantic Imputation to fill occluded regions, (2) Prototype-Guided Tail Mining to strengthen rare-class representations, and (3) EchoOOD Score, a training-free OOD detector fusing local logit coherence with prototype matching.

---

🛠️ Installation	📂 Data	🏋️ Training	📦 Weights	🌿 Branches	📜 Citation

---

🛠️ Installation

Setup details in docs/install.md — Conda, PyTorch 1.9.1 + CUDA 11.1, mmdet3d, spconv, etc.

📂 Data Preparation

Details in docs/dataset.md for SemanticKITTI, KITTI-360, Vaakitti, Vaakitti360, and STU.

🏋️ Training & Evaluation

All commands in docs/run.md. Quick start:

# train (4 GPUs)
./tools/dist_train.sh projects/configs/sgn/proood-semkitti.py 4

# eval
./tools/dist_test.sh projects/configs/sgn/proood-semkitti.py ./path/to/ckpts.pth 4

# OOD
./tools/dist_test_ood.sh projects/configs/sgn/proood-ood-vaakitti.py ./path/to/ckpts.pth 4

📦 Pretrained Weights

Config	Train Set	Train Depth	Test Depth	OOD	Weights
proood-semkitti	SemanticKITTI	MSN	MSN	—	link
proood-sql-semkitti	SemanticKITTI	SQL	SQL	√	link
proood-kitti360	KITTI-360	MSN	MSN	—	link
proood-sql-kitti360	KITTI-360	SQL	SQL	√	link

MSN = MobileStereoNet. All models use the SGN backbone.

🌿 Branches

ProOOD is backbone-agnostic. Two reference implementations are provided:

Branch	Backbone	Status
main	SGN	✅
baseline_b	VoxDet	🚧

git checkout main          # SGN
git checkout baseline_b    # VoxDet

📜 Citation

@article{zhang2026proood,
  title   = {ProOOD: Prototype-Guided Out-of-Distribution 3D Occupancy Prediction},
  author  = {Zhang, Yuheng and Duan, Mengfei and Peng, Kunyu and Wang, Yuhang and
             Wen, Di and Paudel, Danda Pani and Van Gool, Luc and Yang, Kailun},
  journal = {arXiv preprint arXiv:2604.01081},
  year    = {2026}
}

🙏 Acknowledgement

Thanks to the authors of SGN, ProtoOcc, VoxDet, and mmdet3d.