RadarGSOcc: Doppler-Aware 3D Semantic Occupancy Prediction Using 4D Radar Tensor with Geometric-Semantic Feature Extraction

Christofel Rio Goenawan; Dong-Hee Paek; Seung-Hyun Kong

cs.CVOct 13, 2026

RadarGSOcc: Doppler-Aware 3D Semantic Occupancy Prediction Using 4D Radar Tensor with Geometric-Semantic Feature Extraction

Authors: Christofel Rio Goenawan, Dong-Hee Paek, Seung-Hyun Kong

Organizations: Robotics Program, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea · CCS Graduate School of Mobility, KAIST, Daejeon, Korea · Department of Future Mobility, Korea University, Sejong, Korea

Abstract

3D semantic occupancy prediction is crucial for autonomous driving, but camera-based methods degrade in low-light conditions and LiDAR-based methods are sensitive to adverse weather. 4D radar is more robust, yet its sparsity and the high cost of radar tensor processing limit its use. We propose Radar Geometric-Semantic Occupancy Prediction (RadarGSOcc), an efficient framework that directly processes 4D radar tensors by separately extracting geometric and semantic features, including Doppler information, with a deformable attention-based geometric refinement module as its key component. We also introduce the K-Radar 3DSemOcc benchmark, the first large-scale benchmark for 3D semantic occupancy prediction using 4D radar tensors, built on the K-Radar dataset with refined LiDAR-based annotations. RadarGSOcc predicts 15 classes within a 51.2m range and achieves 38.2% IoU, 7.6% mIoU, and 25.5% weighted mIoU, outperforming RadarOcc by 17% relative (25.5 vs. 21.9) in weighted mIoU while using fewer parameters and lower computation. It substantially improves truck prediction while trading off car accuracy, outperforms camera-based approaches, and narrows the gap with LiDAR-based methods. To assess robustness without per-frame ground truth in adverse weather, we further propose a novel proxy metric, Environment-Based KL Divergence (EBKL), under which 4D radar proves a reliable and efficient modality for 3D semantic occupancy prediction. The dataset and code will be publicly available.

Figures & tables

Fig. 1: (Left) Visualization of 3D semantic occupancy prediction under heavy snow in a highway scenario. (a) Input modalities: (left) front-view camera image, (middle) LiDAR point cloud, and (right) BEV of 4D radar tensor. (b) Predicted occupancy using (left) camera-based OpenOcc-C [ 1 ] , (middle) LiDAR-based OpenOcc-L [ 1 ] , and (right) the proposed RadarGSOcc with 4D radar input. RadarGSOcc produces more accurate predictions for key structures such as drivable areas (green box) and highway barriers (red box) under adverse weather. (Right) Quantitative comparison using the proposed EBKL proxy metric (lower is better). RadarGSOcc achieves lower EBKL, indicating more consistent predictions in adverse weather, while camera- and LiDAR-based methods degrade significantly.

Configuration	IoU	mIoU
RadarOcc (with Doppler)	36.4	6.7
RadarOcc (w/o Doppler)	35.9 ( $-$ 0.5)	6.7 (0.0)

TABLE I: Impact of Doppler index on RadarOcc performance (our reproduction on the K-Radar 3DSemOcc benchmark).

Fig. 2: Overview of the K-Radar 3DSemOcc labeling process. (a) Initial LiDAR-based labels. (b) Refined labels using K-Radar annotations, improving moving objects (red box).

Fig. 3: Overview of the proposed RadarGSOcc architecture. The input 4D radar tensor is divided into power and Doppler components and processed by the proposed Geometric-Semantic Feature Extraction module (bottom left, Section III-B2 ), where $W\times H\times L$ denote spherical spatial dimensions and $dim$ is the feature dimension. The extracted features are then passed to the occupancy head (bottom right), where the proposed geometric feature refinement module (Section III-B3 ) refines occupancy features using geometric representations. The final output is a 3D semantic occupancy prediction of size $H\times W\times L$ .

Method	Sensor	IoU	mIoU	W-mIoU	EBKL	DA	Veg	Man	Car	Truck	Ter	Side
CONet-C [ 1 ]	Cam	16.9	3.2	11.3	0.093	27.4	4.2	0.0	0.6	0.0	11.4	0.0
BEVFormer [ 14 ]	Cam	27.3	4.1	16.0	0.072	34.0	10.4	0.1	0.0	0.0	11.9	0.5
OpenOcc-C [ 1 ]	Cam	30.0	5.4	17.1	0.054	34.7	17.3	13.7	5.9	1.8	19.0	0.9
OpenOcc-R [ 1 ]	Radar	26.0	4.0	14.7	0.054	34.0	8.8	5.1	1.2	0.0	8.9	0.8
OpenOcc-RC [ 1 ]	R+Cam	25.2	3.8	14.0	0.075	34.9	5.6	0.8	0.0	0.1	11.0	1.4
4DRC-Occ [ 11 ]	R+Cam	30.2	4.6	17.6	0.054	35.2	12.8	0.1	0.2	0.0	15.3	1.9

TABLE II: 3D semantic occupancy results on the K-Radar 3DSemOcc benchmark (test split). Best, second-best, and third-best results are shown in bold, underlined, and italic, respectively. EBKL is computed under normal weather conditions (lower is better); “–” denotes not evaluated. “+ w/o overlap” denotes RadarGSOcc without the Doppler channels shared between the geometric and semantic branches. RadarOcc and RadarGSOcc rows are the first of three training runs (Section IV-D2 ). Results for all classes are given in Table S-VI of the supplementary material.

Fig. 4: Visualization of 3D semantic occupancy prediction using LiDAR (OpenOcc-L) [ 1 ] , camera (OpenOcc-C) [ 1 ] , and 4D radar tensors with RadarOcc [ 7 ] and the proposed RadarGSOcc in urban (1) and highway (2) scenarios. In each scenario, the top row shows (from left) the front-camera image, LiDAR point cloud, 4D radar tensor, and ground truth, while the bottom row shows camera-based prediction, LiDAR-based prediction, RadarOcc, and RadarGSOcc. Each prediction includes a 3D front-right view, BEV view, and front view. In these examples, RadarGSOcc predicts the highlighted moving objects (red and black boxes) more accurately, improves terrain and drivable area prediction (dashed black box), and reduces false moving-object predictions (dashed red box); quantitatively, however, its merged vehicle IoU is lower than that of RadarOcc (Section IV-D1 ).

Threshold	mIoU	W-mIoU	#Voxel GT	#Voxel Prediction
0.2	3.97	12.7	26,647	14,842
0.4	5.44	17.7	25,796	17,130
0.8	7.60	25.5	14,606	9,930
0.9	7.14	24.5	10,559	7,049

TABLE III: Performance of RadarGSOcc on the K-Radar 3DSemOcc benchmark under different class probability thresholds.

Fusion	Conv	Def	IoU	mIoU	W-mIoU	Param	FLOPs	FPS
RadarOcc [ 7 ]			36.4	6.7	21.9	117.5	1334.1	3.3
$\times$	$\times$	$\times$	35.7	6.9	21.8	104.2	1260.1	3.2
$\times$	$\times$	$\checkmark$	36.1	7.0	21.1	104.6	1270.8	3.1
$\checkmark$	$\times$	$\times$	35.9	6.6	20.7	104.2	1262.7	3.1
$\checkmark$	$\checkmark$	$\times$	32.3	5.6	18.1	105.0	1285.4	3.2
$\checkmark$	$\times$	$\checkmark$	38.2	7.6	25.5	104.6	1273.5	3.1

TABLE IV: Ablation study of RadarGSOcc. Fusion: geometric-semantic fusion; Conv/Def: geometric refinement with 3D convolution / deformable attention.

\checkmark

indicates enabled. Param in millions (M) and FLOPs in GFLOPs (G). Each variant is a single training run.

Fig. 5: Visualization of 3D occupancy features before (left) and after (middle) geometric refinement, compared with ground truth (right). In these examples, the refined features respond more strongly to moving objects (cars and trucks in red and black boxes) and road structures (dashed red boxes).

Top-K	Doppler	IoU	mIoU	W-mIoU	FLOPs	FPS
150	1	35.2	6.3	20.0	1273.2	3.2
150	2	35.8	6.8	21.2	1273.2	3.2
150	3	36.5	6.8	20.4	1273.2	3.2
250	1	36.9	6.6	21.5	1273.3	3.1
250	2	34.5	6.8	20.2	1273.4	3.1
250	3	38.2	7.6	25.5	1273.5	3.1

TABLE V: Ablation of RadarGSOcc with different Top-K values and Doppler dimensions (single training runs).

Fig. 6: mIoU versus EBKL (normal weather, test split) of the nine models in Table II for which EBKL is evaluated. The dashed line is a least-squares fit; the values are listed in Table S-IX of the supplementary material.

Model	Normal	Sleet	Rain	Fog	Light	Heavy
OpenOcc-C [ 1 ]	18.5	17.5	16.7	16.5	17.0	18.9
CONet-L [ 1 ]	29.8	9.7	14.3	9.5	9.8	14.2
RadarGSOcc	26.5	26.1	24.5	28.1	15.1	15.3

TABLE VI: Percentage of non-empty voxel predictions under different weather conditions (Light/Heavy: light/heavy snow). The adverse-weather sequences are listed in Table S-I of the supplementary material.

Scene	Split	Weather	Environment	Time	#Frames
2	Train	Normal	Highway	Night	462
5	Train	Normal	Urban	Day	595
8	Train	Normal	University	Night	567
9	Train	Normal	Highway	Day	833
10	Train	Normal	Highway	Day	1,130
14	Train	Normal	Urban	Day	595

TABLE S-I: Composition of the K-Radar 3DSemOcc benchmark and the unlabeled adverse-weather sequences.

Class	#Voxels (train)	%	#Voxels (test)	%
Drivable area	4,183,019	24.12	1,216,791	26.67
Vegetation	8,536,616	49.22	2,068,116	45.34
Manmade	1,776,829	10.24	382,270	8.38
Terrain	2,056,773	11.86	710,776	15.58
Sidewalk	406,188	2.34	91,606	2.01
Other flat	12,819	0.07	6,723	0.15

TABLE S-II: Ground-truth voxel statistics per class before label refinement (label-confidence threshold 0.8).

Class	#Voxels (train)	%	#Voxels (test)	%
Drivable area	3,738,578	22.94	1,206,770	25.95
Vegetation	7,977,664	48.96	2,048,675	44.06
Manmade	1,526,764	9.37	372,645	8.01
Terrain	1,888,009	11.59	704,612	15.15
Sidewalk	378,405	2.32	91,151	1.96
Other flat	12,065	0.07	6,518	0.14

TABLE S-III: Ground-truth voxel statistics per class after label refinement (label-confidence threshold 0.8), used in all experiments.

Fig. S1: Manual verification of the K-Radar 3DSemOcc benchmark. Sampled voxels (green 3D boxes) are projected onto the front-camera image (left) and the LiDAR bird’s-eye view (right): (top) car voxels in a night scene and (bottom) manmade voxels in a day scene.

Class	#Sampled	#Correct	Accuracy (%)
Drivable area	126	125	99.2
Vegetation	191	188	98.4
Manmade	66	64	97.0
Terrain	99	69	69.7
Car	16	14	87.5
Truck/bus	5	4	80.0

TABLE S-IV: Manual verification of 525 randomly sampled non-empty voxels.

	Self-attn.	Fusion	Refinement
Feature dimension $dim$	96	96	192
Number of heads	8	8	8
Sampling points per query	8	8	8
Number of layers	1	1	1
Sampling coordinates	$[0,1]$	$[0,1]$	$[0,1]$

TABLE S-V: Hyperparameters of the deformable attention modules.

Method	Sensor	IoU	mIoU	W-mIoU	DA	Veg	Man	Car	Truck	Ter	Side	Other	Trail	Cone	Ped	Moto	Bike	Bar
CONet-C [ 1 ]	Cam	16.9	3.2	11.3	27.4	4.2	0.0	0.6	0.0	11.4	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
BEVFormer [ 14 ]	Cam	27.3	4.1	16.0	34.0	10.4	0.1	0.0	0.0	11.9	0.5	0.0	0.0	0.0	0.0	0.0	0.0	0.0
OpenOcc-C [ 1 ]	Cam	30.0	5.4	17.1	34.7	17.3	13.7	5.9	1.8	19.0	0.9	1.6	0.0	0.0	0.0	0.0	0.0	1.5
OpenOcc-R [ 1 ]	Radar	26.0	4.0	14.7	34.0	8.8	5.1	1.2	0.0	8.9	0.8	0.0	0.0	0.0	0.0	0.0	0.0	0.0
OpenOcc-RC [ 1 ]	R+Cam	25.2	3.8	14.0	34.9	5.6	0.8	0.0	0.1	11.0	1.4	0.0	0.0	0.0	0.0	0.0	0.0	0.0
4DRC-Occ [ 11 ]	R+Cam	30.2	4.6	17.6	35.2	12.8	0.1	0.2	0.0	15.3	1.9	0.0	0.0	0.0	0.0	0.0	0.0	0.0

TABLE S-VI: 3D semantic occupancy results on the K-Radar 3DSemOcc benchmark for all classes (IoU in %, test split).

Fig. S2: Normalized ground-truth class distributions (%) of the normal-weather scenes, grouped by environment, for the eight classes shown.

Scenes $\backslash$ Reference	Urban	University	Highway
Urban scenes	0.017	0.014	0.030
University scenes	0.015	0.008	0.035
Highway scenes	0.032	0.036	0.015

TABLE S-VII: Mean EBKL between scene-level and environment-level class distributions (ground truth, normal weather).

Model	Run (seed)	IoU	mIoU	W-mIoU
RadarOcc	1 (not fixed)	36.4	6.7	21.9
	2 (1111)	36.3	6.4	20.1
	3 (2222)	36.0	7.0	20.8
	Mean $\pm$ std	36.26 $\pm$ 0.23	6.72 $\pm$ 0.28	20.93 $\pm$ 0.90
RadarGSOcc	1 (not fixed)	38.2	7.6	25.5
	2 (1111)	39.3	7.0	24.6

TABLE S-VIII: Results of the three training runs of RadarOcc and RadarGSOcc on the K-Radar 3DSemOcc benchmark (test split).

Model	mIoU (%)	EBKL
CONet-C [ 1 ]	3.2	0.093
BEVFormer [ 14 ]	4.1	0.072
OpenOcc-C [ 1 ]	5.4	0.0536
OpenOcc-R [ 1 ]	4.0	0.054
OpenOcc-RC [ 1 ]	3.8	0.075
4DRC-Occ [ 11 ]	4.6	0.0543

TABLE S-IX: mIoU and EBKL (normal weather, test split) of the nine models in Fig. 6 .

3D Occupancy Prediction

3D occupancy prediction estimates the geometry and semantics of occupied and free space around an agent from camera, LiDAR, or radar data.

See all 113 papers

4DRC-OCC: Robust Semantic Occupancy Prediction Through Fusion of 4D Radar and Camera
Mar 8, 2026David Ninfa, Andras Palffy, Holger Caesar3D Occupancy Prediction Autonomous Driving Perception
Cognitive Robotics, Delft University of Technology · Perciv AI
4DR360: State Reasoning for Joint 3D Detection and Occupancy Prediction in 4D Radar-Camera Full-Scene Perception
Jul 10, 2026Xiaokai Bai, Lianqing Zheng, Runwei Guan +3Camera-Radar Fusion Millimeter-Wave Radar
College of Information Science and Electronic Engineering, Zhejiang University. · School of Automotive Studies, Tongji University. · Thrust of Artificial Intelligence, Hong Kong University of Science and Technology.
VLMFusionOcc3D: VLM Assisted Multi-Modal 3D Semantic Occupancy Prediction
Mar 3, 2026A. Enes Doruk, Hasan F. Ates3D Occupancy Prediction Autonomous Driving Perception

Mar 8, 2026J/K move · Enter open · S save

4DRC-OCC: Robust Semantic Occupancy Prediction Through Fusion of 4D Radar and Camera

David Ninfa, Andras Palffy, Holger Caesar

Cognitive Robotics, Delft University of Technology · Perciv AI

3D Occupancy Prediction Autonomous Driving Perception Camera-Radar Fusion Autonomous Driving Datasets

Open paper AskPDF arXiv

Autonomous driving requires robust perception across diverse environmental conditions, yet 3D semantic occupancy prediction remains challenging under adverse weather and lighting. In this work, we present the first study combining 4D radar and camera data for 3D semantic occupancy prediction. Our fusion leverages the complementary strengths of both modalities: 4D radar provides reliable range, velocity, and angle measurements in challenging conditions, while cameras contribute rich semantic and texture information. We further show that integrating depth cues from camera pixels enables lifting 2D images to 3D, improving scene reconstruction accuracy. Additionally, we introduce a fully automatically labeled dataset for training semantic occupancy models, substantially reducing reliance on costly manual annotation. Experiments demonstrate the robustness of 4D radar across diverse scenarios, highlighting its potential to advance autonomous vehicle perception.

Abstract

3D Occupancy Prediction

Explore similar work

4DRC-OCC: Robust Semantic Occupancy Prediction Through Fusion of 4D Radar and Camera

4DR360: State Reasoning for Joint 3D Detection and Occupancy Prediction in 4D Radar-Camera Full-Scene Perception

VLMFusionOcc3D: VLM Assisted Multi-Modal 3D Semantic Occupancy Prediction