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

Yuheng Zhang; Mengfei Duan; Kunyu Peng; Yuhang Wang; Di Wen; Danda Pani Paudel; Luc Van Gool; Kailun Yang

cs.CVApr 1, 2026

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

Authors: Yuheng Zhang, Mengfei Duan, Kunyu Peng, Yuhang Wang, Di Wen, Danda Pani Paudel, Luc Van Gool, Kailun Yang

Organizations: Hunan University · Karlsruhe Institute of Technology · INSAIT, Sofia University “St. Kliment Ohridski”

Abstract

3D semantic occupancy prediction is central to autonomous driving, yet current methods are vulnerable to long-tailed class bias and out-of-distribution (OOD) inputs, often overconfidently assigning anomalies to rare classes. We present ProOOD, a lightweight, plug-and-play method that couples prototype-guided refinement with training-free OOD scoring. ProOOD comprises (i) prototype-guided semantic imputation that fills occluded regions with class-consistent features, (ii) prototype-guided tail mining that strengthens rare-class representations to curb OOD absorption, and (iii) EchoOOD, which fuses local logit coherence with local and global prototype matching to produce reliable voxel-level OOD scores. Extensive experiments on five datasets demonstrate that ProOOD achieves state-of-the-art performance on both in-distribution 3D occupancy prediction and OOD detection. On SemanticKITTI, it surpasses baselines by +3.57% mIoU overall and +24.80% tail-class mIoU; on VAA-KITTI, it improves AuPRCr by +19.34 points, with consistent gains across benchmarks. These improvements yield more calibrated occupancy estimates and more reliable OOD detection in safety-critical urban driving. The source code is publicly available at https://github.com/7uHeng/ProOOD.

Figures & tables

Figure 1: Overview of the proposed ProOOD framework. Coarse 3D features are derived from a 2D backbone and a view transformation module. These features are then refined by PGSI and PGTM modules, which leverage class-wise prototypes updated via EMA to enhance semantic completion and improve tail-class sensitivity. The resulting refined features are used to update the class prototypes, which in turn support both semantic occupancy prediction and OOD detection through the EchoOOD mechanism.

Figure 2: (a) Generation of Local and Global Voxel Prototypes. (b) EchoOOD. It computes OOD maps via three cues: local logit alignment, and local/global prototype matching.

Method	IoU	road (15.30%)	sidewalk (11.13%)	parking (1.12%)	other-grnd. (0.56%)	building (14.1%)	car (3.92%)	truck (0.16%)	bicycle (0.03%)	motorcycle (0.03%)	other-veh. (0.20%)	vegetation (39.3%)	trunk (0.51%)	terrain (9.17%)	person (0.07%)	bicyclist (0.07%)	motorcyclist (0.05%)	fence (3.90%)	pole (0.29%)	traf.-sign (0.08%)	mIoU	Tail mIoU
SemanticKITTI
MonoScene [ 7 ]	34.16	54.70	27.10	24.80	5.70	14.40	18.80	3.30	0.50	0.70	4.40	14.90	2.40	19.50	1.00	1.40	0.40	11.10	3.30	2.10	11.08	2.29
TPVFormer [ 31 ]	34.25	55.10	27.20	27.40	6.50	14.80	19.20	3.70	1.00	0.50	2.30	13.90	2.60	20.40	1.10	2.40	0.30	11.00	2.90	1.50	11.26	2.25
OccFormer [ 107 ]	34.53	55.90	30.30	31.50	6.50	15.70	21.60	1.20	1.50	1.70	3.20	16.80	3.90	21.30	2.20	1.10	0.20	11.90	3.80	3.70	12.32	2.64
VoxFormer [ 52 ]	42.95	53.90	25.30	21.10	5.60	19.80	20.80	3.50	1.00	0.70	3.70	22.40	7.50	21.30	1.40	2.60	0.20	11.10	5.10	4.90	12.20	3.29
HASSC [ 89 ]	43.40	54.60	27.70	23.80	6.20	21.10	22.80	4.70	1.60	1.00	3.90	23.80	8.50	23.30	1.60	4.00	0.30	13.10	5.80	5.50	13.34	3.92

Table 1: Quantitative results on the SemanticKITTI test set [ 4 ] and VAA-KITTI [ 106 ] (Single-Frame Evaluation for Fair Comparison). * denotes results reproduced using the official code implementation.

		VAA-KITTI				VAA-KITTI-360				VAA-STU
Methods	Anomaly Score Methods	AuPRC r $\uparrow$			AuROC $\uparrow$	AuPRC r $\uparrow$			AuROC $\uparrow$	AuPRC r $\uparrow$			AuROC $\uparrow$
Methods	Anomaly Score Methods	$0.8m$	$1.0m$	$1.2m$	AuROC $\uparrow$	$0.8m$	$1.0m$	$1.2m$	AuROC $\uparrow$	$0.8m$	$1.0m$	$1.2m$	AuROC $\uparrow$
VoxDet [ 49 ]	ASS [ 106 ]	4.57	8.36	14.02	59.52	11.79	25.08	48.49	50.76	1.45	3.05	5.89	59.52
SGN [ 63 ]		8.22	14.45	22.93	60.51	13.27	28.57	52.99	47.27	2.91	5.74	11.32	64.51
OccOoD [ 106 ]		10.80	16.05	23.42	61.96	14.26	29.72	54.35	47.84	3.05	6.15	11.39	65.37
Ours (+VoxDet)	EchoOOD	21.95	36.97	56.49	60.99	12.75	26.81	50.08	47.24	2.74	5.09	8.82	55.11

Table 2: Quantitative OOD results on VAA-KITTI, VAA-KITTI-360, and VAA-STU datasets [ 106 ] .

Method	IoU	car (2.85%)	bicycle (0.01%)	motorcycle (0.01%)	truck (0.16%)	other-veh. (5.75%)	person (0.02%)	road (14.98%)	parking (2.31%)	sidewalk (6.43%)	other-grnd. (2.05%)	building (15.67%)	fence (0.96%)	vegetation (41.99%)	terrain (7.10%)	pole (0.22%)	traf.sign (0.06%)	other-struct. (4.33%)	other-obj. (0.28%)	mIoU	Tail mIoU
SSCBench-KITTI-360
MonoScene [ 7 ]	37.87	19.34	0.43	0.58	8.02	2.03	0.86	48.35	11.38	28.13	3.32	32.89	3.53	26.15	16.75	6.92	5.67	4.20	3.09	12.31	3.61
TPVFormer [ 31 ]	40.22	21.56	1.09	1.37	8.06	2.57	2.38	52.99	11.99	31.07	3.78	34.83	4.80	30.08	17.52	7.46	5.86	5.48	2.70	13.64	4.13
VoxFormer [ 52 ]	38.76	17.84	1.16	0.89	4.56	2.06	1.63	47.01	9.67	27.21	2.89	31.18	4.97	28.99	14.69	6.51	6.92	3.79	2.43	11.91	3.44
OccFormer [ 107 ]	40.27	22.58	0.66	0.26	9.89	3.82	2.77	54.30	13.44	31.53	3.55	36.42	4.80	31.00	19.51	7.77	8.51	6.95	4.60	13.81	4.92
Symphonies [ 35 ]	44.12	30.02	1.85	5.90	25.07	12.06	8.20	54.94	13.83	32.76	6.93	35.11	8.58	38.33	11.52	14.01	9.57	14.44	11.28	18.58	10.84

Table 3: Quantitative results on the SSCBench-KITTI-360 test set [ 51 ] and VAA-KITTI-360 [ 106 ] (Single-Frame Evaluation for Fair Comparison). * denotes results reproduced using the official code implementation.

Baseline	PBCL	PGSI	PGTM	IoU	mIoU	Tail mIoU
✓				43.60	14.55	4.07
✓	✓			43.69	14.63	4.44 (+9.09%)
✓	✓	✓		44.09	14.76	4.58 (+3.15%)
✓	✓	✓	✓	44.02	14.97	4.89 (+6.77%)

Table 4: Ablation study of each component on the SemanticKITTI [ 4 ] validation set for 3D Occupancy prediction.

Anomaly Score Methods	VAA-KITTI
	AuPRC r $\uparrow$			AuROC $\uparrow$
	$0.8m$	$1.0m$	$1.2m$	AuROC $\uparrow$
Postpro	5.81	10.02	16.22	63.85
Energy	6.80	12.11	20.08	66.69
Entropy	7.38	12.91	21.05	66.24
ASS [ 106 ]	14.47	20.96	29.65	64.20

Table 5: Ablation studies of anomaly score methods on the VAA-KITTI dataset [ 106 ] for OccOoD evaluation.

Figure 3: Qualitative results of 3D occupancy prediction on the SemanticKITTI validation set [ 4 ] .

Figure 4: Qualitative results of out-of-distribution detection on the 07VAA sequence of VAA-KITTI [ 106 ] .

Method	road (15.30%)	sidewalk (11.13%)	parking (1.12%)	other-grnd. (0.56%)	building (14.1%)	car (3.92%)	truck (0.16%)	bicycle (0.03%)	motorcycle (0.03%)	other-veh. (0.20%)	vegetation (39.3%)	trunk (0.51%)	terrain (9.17%)	person (0.07%)	bicyclist (0.07%)	motorcyclist (0.05%)	fence (3.90%)	pole (0.29%)	traf.-sign (0.08%)	mIoU	Tail mIoU
SemanticKITTI
MonoScene [ 7 ]	57.47	27.05	15.72	0.87	14.24	23.55	7.83	0.20	0.77	3.59	18.12	2.57	30.76	1.79	1.03	0.00	6.39	4.11	2.48	11.50	2.29
TPVFormer [ 31 ]	56.50	25.87	20.60	0.85	13.88	23.81	8.08	0.36	0.05	4.35	16.92	2.26	30.38	0.51	0.89	0.00	5.94	3.14	1.52	11.36	2.00
VoxFormer [ 52 ]	54.76	26.35	15.50	0.70	17.65	25.79	5.63	0.59	0.51	3.77	24.39	5.08	29.96	1.78	3.32	0.00	7.64	7.11	4.18	12.35	2.97
OccFormer [ 107 ]	58.85	26.88	19.61	0.31	14.40	25.09	25.53	0.81	1.19	8.52	19.63	3.93	32.62	2.78	2.82	0.00	5.61	4.26	2.86	13.46	4.82
NDC-scene [ 100 ]	59.20	28.24	21.42	1.67	14.94	26.26	14.75	1.67	2.37	7.73	19.09	3.51	31.04	3.60	2.74	0.00	6.65	4.53	2.73	12.70	4.12

Table 6: Quantitative results on the SemanticKITTI val set [ 4 ] (single-frame evaluation for fair comparison). * denotes results reproduced using the official code implementation.

Methods	$ECE_{sem}$ $\downarrow$	$ECE_{tail}$ $\downarrow$	$ECE_{geo}$ $\downarrow$	mIoU $\uparrow$	Tail mIoU $\uparrow$	AuROC $\uparrow$
SGN	51.10	57.49	4.93	11.10	2.20	60.51
Ours (+SGN)	47.28	53.91	4.41	11.86	2.76	64.31

Table 7: Quantitative evaluation of calibration quality.

EMA Momentum	mIoU	Tail mIoU
0.01	14.75	4.32
0.05	14.97	4.89
0.10	15.14	4.55
0.20	15.01	4.36

Table 8: Ablation study on the impact of different EMA momentum parameters for 3D occupancy prediction on the SemanticKITTI validation set [ 4 ] .

Exp. ID	$\eta$	$\delta$	mIoU	Tail mIoU
1	0.1	0.10	14.89	4.60
2	0.3	0.05	14.75	4.25
3	0.3	0.10	14.97	4.89
4	0.3	0.20	15.16	4.70
5	0.5	0.10	14.84	4.38

Table 9: Ablation study of hyperparameters.

Warm-up Iterations	mIoU	Tail mIoU
500	15.08	4.68
750	14.97	4.89
1000	15.09	4.80
1500	15.01	4.73

Table 10: Ablation on different prototype warm-up iterations for 3D occupancy prediction on the SemanticKITTI validation set [ 4 ] .

Class	Percentage	Tail
road	15.30%	–
sidewalk	11.13%	–
parking	1.12%	–
other-ground	0.56%	✓
building	14.10%	–
car	3.92%	–

Table 11: Class distribution in SemanticKITTI [ 4 ] . Tail classes (proportion

<0.60\%

) are marked with ✓.

Class	Percentage	Tail
car	2.85%	–
bicycle	0.01%	✓
motorcycle	0.01%	✓
truck	0.16%	✓
other-vehicle	5.75%	–
person	0.02%	✓

Table 12: Class distribution in SSCBench-KITTI-360. Tail classes (proportion

<0.30\%

) are marked with ✓.

Methods	mIoU	Tail mIoU
VoxFormer	12.86	2.87
Ours (+VoxFormer)	13.32 (+0.46)	3.15 (+0.28)
CGFormer	15.64	4.18
Ours (+CGFormer)	16.24 (+0.60)	4.62 (+0.44)

Table 13: Plug-and-play performance.

Method	SGN [ 63 ]	ProOOD (+SGN)	VoxDet [ 49 ]	ProOOD (+VoxDet)
#Parameter (M)	28.16	28.44	55.34	55.35
FLOPs (G)	522.34	580.66	1311.10	1311.48

Table 14: Comparison of model efficiency with baselines.

Methods	FPS	Inf. Mem.	mIoU	Tail mIoU
SGN	7.6	5.16 (GB)	14.55	4.07
Ours (+SGN)	7.0	5.30 (GB)	14.97	4.89
VoxDet	5.0	4.34 (GB)	18.08	7.44
Ours (+VoxDet)	4.9	4.47 (GB)	18.86	8.46

Table 15: FPS and GPU memory comparisons.

Figure 5: Failure case due to inaccurate occupancy prediction for OOD voxels.

Figure 6: Failure cases under challenging long-range and heavy occlusion scenarios.

Explore similar work

SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction
Jul 6, 2026Pin Tang, Zhongdao Wang, Guoqing Wang +23D Reconstruction Autonomous Driving Perception
MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai 200240, China · Huawei Noah’s Ark Lab, China
Class-Distribution Guided Active Learning for 3D Occupancy Prediction in Autonomous Driving
Mar 28, 2026Wonjune Kim, In-Jae Lee, Sihwan Hwang +2Active Learning 3D Occupancy Prediction
Robotics Program, KAIST, Daejeon 34141, South Korea · Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul 08826, South Korea · Graduate School of Mobility, KAIST, Daejeon 34051, South Korea +1
$M^2$ -Occ: Resilient 3D Semantic Occupancy Prediction for Autonomous Driving with Incomplete Camera Inputs
Mar 10, 2026Kaixin Lin, Kunyu Peng, Di Wen +3Multi-View Learning Autonomous Driving Perception
School of Artificial Intelligence and Robotics and the National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, China · Institute for Anthropomatics and Robotics, Karlsruhe Institute of Technology, Germany

Jul 6, 2026J/K move · Enter open · S save

SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction

Pin Tang, Zhongdao Wang, Guoqing Wang, Xiangxuan Ren+1

MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai 200240, China · Huawei Noah’s Ark Lab, China

3D Reconstruction Autonomous Driving Perception 3D Semantic Segmentation Semantic Occupancy Mapping 3D Occupancy Prediction 3D Scene Completion

Open paper AskPDF arXiv

Vision-based 3D semantic occupancy prediction is essential for autonomous driving, yet dense voxel representations waste computation on largely empty space, while BEV and TPV projections compromise fine-grained 3D structure. Fully sparse representations offer an attractive alternative, but existing methods, including SparseOcc, entangle scene completion with semantic prediction by indiscriminately propagating high-dimensional features into empty regions and applying voxel-wise classification. This creates excessive activations, computational overhead, and geometric ambiguity. We present SparseOcc++, a geometry-aware sparse framework that explicitly decouples scene completion from semantic segmentation. SparseOcc++ reformulates completion as signed-distance regression on sparse anchor voxels through a scene completion field (SCF). To model complex outdoor geometry robustly, it combines orthogonal decomposition with discretized distance learning. A geometry-guided propagation module then converts the SCF into a complete volumetric scene and restricts semantic segmentation to geometrically verified regions. Experiments establish new state of the art: SparseOcc++ improves IoU by 2.3 points and is 3.9x faster than SparseOcc on nuScenes, while achieving a 5.9x speedup over OccFormer on SemanticKITTI.

Abstract

Explore similar work

SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction

Class-Distribution Guided Active Learning for 3D Occupancy Prediction in Autonomous Driving

M2M^2M2-Occ: Resilient 3D Semantic Occupancy Prediction for Autonomous Driving with Incomplete Camera Inputs

$M^2$ -Occ: Resilient 3D Semantic Occupancy Prediction for Autonomous Driving with Incomplete Camera Inputs