DualPathOcc: Dual-Resolution BEV Encoder for 3D Occupancy Prediction

Lihao Qiu; Jian Chen; Ruihao Wang; Ramu Gautam; Mei Yang; Yingtao Jiang

cs.CVSep 6, 2026

DualPathOcc: Dual-Resolution BEV Encoder for 3D Occupancy Prediction

Authors: Lihao Qiu, Jian Chen, Ruihao Wang, Ramu Gautam, Mei Yang, Yingtao Jiang

Organizations: Department of Electrical and Computer Engineering, University of Nevada, Las Vegas · Institute of Logistics Science and Engineering, Shanghai Maritime University · School of Software Engineering, Xi’an Jiaotong University

Abstract

Predicting 3D occupancy from multi-view images requires preserving geometric detail during 2D-to-3D lifting while reasoning over sparse, volumetric scene representations. We present DualPathOcc, a camera-based framework that combines a Spatial Enhancer for high-resolution feature aggregation before BEV compression, a SENet-augmented dual-path BEV encoder for local-global context modeling, and height-aware weighted cross-entropy for near-ground occupancy. The final model is optimized with occupancy supervision and no explicit depth loss. On single-frame Occ3D-nuScenes, DualPathOcc achieves 37.37 mIoU. We further analyze how surface-centered depth targets interact with volumetric occupancy learning.

Figures & tables

Figure 1: Overview of DualPathOcc. Multi-view features are lifted by LSS, refined by the Spatial Enhancer before BEV compression, and processed by local and global encoder paths with channel recalibration before occupancy prediction. $\Sigma$ denotes element-wise addition and $\oplus$ denotes channel concatenation. The final model uses no explicit depth loss; depth-target variants are analyzed separately in Section A .

Figure 2: Spatial Enhancer. After folding four height bins into channels, two 2D convolution–downsampling stages transform $B\times(4C)\times 800\times 800$ lifted features into the $B\times(4C)\times 200\times 200$ representation used by the dual-path BEV encoder.

Method	others	barriers	bicycle	bus	car	Const. veh.	motorcycle	pedestrian	traffic cone	trailer	truck	Dri. Sur.	other flat	sidewalk	terrain	manmade	vegetation	mIoU
MonoScene [ 2 ]	1.7	7.2	4.2	4.9	9.3	5.6	3.9	3.0	5.9	4.4	7.1	14.9	6.3	7.9	7.4	1.0	7.6	6.0
OccFormer [ 37 ]	5.9	30.2	12.3	34.4	39.1	14.4	16.4	17.2	9.2	13.9	26.3	50.9	30.9	34.6	22.7	6.7	6.9	21.9
TPVFormer [ 9 ]	7.2	38.9	13.6	40.7	45.9	17.2	19.9	18.8	14.3	26.6	34.1	55.6	35.4	37.5	30.7	19.4	16.7	27.8
CTF-Occ [ 25 ]	8.09	39.33	20.56	38.29	42.24	16.93	24.52	22.72	21.05	22.98	31.11	53.33	33.84	37.98	33.23	20.79	18.0	28.53
RenderOcc [ 23 ]	4.84	31.72	10.72	27.67	26.45	13.87	18.2	17.67	17.84	21.19	23.25	63.2	36.42	46.21	44.26	19.58	20.72	26.11
PanoOcc [ 29 ]	8.6	43.7	21.6	42.5	49.9	21.3	25.3	22.9	20.1	29.7	37.1	80.9	40.3	49.6	52.8	39.8	35.8	36.6

Table 1: 3D occupancy prediction performance on the Occ3D-nuScenes validation set. “Const. Veh.” and “Dri. Sur.” denote construction vehicle and drivable surface, respectively.

Figure 3: Illustrative DualPathOcc predictions on Occ3D-nuScenes. Top: front and front-right camera views. Bottom: the corresponding semantic occupancy prediction. Voxels follow the class colors in Table 1 , and outlines associate objects between image and occupancy views.

Configuration	mIoU
FlashOcc baseline	32.08
Baseline + Spatial Enhancer	35.31
Baseline + height-aware loss	35.71
Baseline + dual-path BEV encoder	36.20
Baseline + dual-path BEV encoder + SENet	36.92
Encoder + SENet + Spatial Enhancer	37.15

Table 2: Component configurations on Occ3D-nuScenes. All variants share the

256\times 704

image input, batch size, AdamW optimizer, 24-epoch schedule, and no explicit depth loss. The variants are not capacity-matched because the listed modules add capacity; the Spatial Enhancer changes the intermediate grid from

200\times 200

800\times 800

before returning to

200\times 200

Model	Image Backbone	Image Size	mIoU
BEVFormer	ResNet-101	$928\times 1600$	26.88
FastOcc	ResNet-101	$928\times 1600$	39.21
PanoOcc	ResNet-101	$928\times 1600$	42.13
BEVDet4D	ResNet-50	$384\times 704$	39.25
FB-OCC	ResNet-50	$256\times 704$	40.69
FlashOcc	ResNet-50	$256\times 704$	37.84

Table 3: Comparison with multi-frame methods on the Occ3D-nuScenes validation set. DualPathOcc variants use one aligned historical frame. The base model uses a

400\times 400

Spatial Enhancer grid and batch size 4; * uses

800\times 800

and batch size 2. The ~ and ^ variants use

512\times 1408

images with

400\times 400

and

800\times 800

grids, respectively, and batch size 2.

Depth supervision method	mIoU
One-hot	35.96
No supervision	36.92
Gaussian ( $\sigma=1.5$ )	36.94

Table 4: Comparison of models trained with different depth supervision strategies on the Occ3D-nuScenes dataset. All models in this experiment exclude the Spatial Enhancer module.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Comparison of ground-truth depth representations on a sample image. (a) One-hot encoded depth distribution. (b) Gaussian-encoded depth distribution. $H$ and $W$ denote image height and width. Each depth bin corresponds to the distance of a pixel projected into 3D space relative to the ego vehicle. In the one-hot encoding, each pixel activates only a single bin with a value of 1, while in the Gaussian encoding, the area under the Gaussian curve for each pixel sums to 1, providing a smoother depth representation.

Panel of Figure 5 — Figure 5: Qualitative comparison of depth supervision strategies. (a) Original camera input showing two nearby vehicles. (b) One-hot depth supervision, where green pixels indicate voxels classified at the 7 m distance bin. (c) Model trained without depth supervision, predicting the 7 m region with weaker spatial focus (blue). (d) Gaussian-encoded depth supervision, showing an intermediate spread of responses in the 7 m bin (cyan).

Panel of Figure 6 — Figure 6: Comparison of BEV feature projections under different depth supervision strategies. (a) One-hot encoded depth supervision, (b) no depth supervision, and (c) Gaussian-encoded depth supervision. Gray regions indicate undefined areas, while white and black denote high and low feature responses, respectively.

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

LightOcc: Lightweight Spatial Embedding for Efficient Vision-based 3D Occupancy Prediction
Dec 8, 2024Jinqing Zhang, Yanan Zhang, Wenrui Cai +2Bird's-Eye-View Representation 3D Occupancy Prediction
State Key Laboratory of Virtual Reality Technology and Systems, Beihang University. · School of Computer Science and Information Engineering, Hefei University of Technology. · Hangzhou Innovation Institute, Beihang University.
VGGT-Occ: Geometry-Grounded and Density-Aware Gated Fusion for 3D Occupancy Prediction
May 16, 2026Xun Chen, Tianchen Deng, Rui Wang +53D Semantic Segmentation 3D Occupancy Prediction
Nanyang Technological University · Shanghai Jiao Tong University · ETH Zurich
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

Dec 8, 2024J/K move · Enter open · S save

LightOcc: Lightweight Spatial Embedding for Efficient Vision-based 3D Occupancy Prediction

Jinqing Zhang, Yanan Zhang, Wenrui Cai, Qingjie Liu+1

State Key Laboratory of Virtual Reality Technology and Systems, Beihang University. · School of Computer Science and Information Engineering, Hefei University of Technology. · Hangzhou Innovation Institute, Beihang University.

Bird's-Eye-View Representation 3D Occupancy Prediction

Open paper AskPDF arXiv

Occupancy prediction has garnered increasing attention in recent years for its comprehensive fine-grained environmental representation and strong generalization to open-set objects. Nevertheless, mainstream occupancy prediction methods employ cumbersome voxel features as the scene representation, incurring substantial overheads in both memory and computation. When comparing the occupancy distribution in each spatial dimension, we find that the information entropy of the height dimension is much lower than the other two dimensions that constitute the Bird's Eye View (BEV) plane, which indicates that the height distribution of occupancy is easier to learn and predict. Accordingly, we propose Lightweight Spatial Embedding that can represent complete height information in a more compact way than voxel features, thus significantly enhancing its deployability. First, Single-Channel Occupancy is sampled from the multi-view depth distributions, which is then processed by Spatial-to-Channel mechanism to extract Lightweight Spatial Embeddings of different views by 2D convolution. These embeddings will interact with each other through the Lightweight Cross-View Interaction module to obtain the Unified Embedding, which can directly supplement BEV features with height information. Furthermore, we extract Edge-aware Spatial Embedding and apply Geometric Supervision on Spatial Embeddings, aiming to enhance their capability to represent spatial information. We also propose BEV-CutMix, a feature-level data augmentation strategy, to increase the diversity of the driving scenes. We integrate these innovative components into a pure 2D convolutional model, namely LightOcc. Sufficient experimental results show that LightOcc achieves state-of-the-art performance on multiple benchmarks while demonstrating significant efficiency advantages.

Abstract

3D Occupancy Prediction

Explore similar work

LightOcc: Lightweight Spatial Embedding for Efficient Vision-based 3D Occupancy Prediction

VGGT-Occ: Geometry-Grounded and Density-Aware Gated Fusion for 3D Occupancy Prediction

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