_{Representative InfraOcc keyframe. Synchronized roadside views are associated with dense semantic occupancy in a fixed infrastructure coordinate system.}

🗓️ News

• 2026.07.01: Paper submitted.
• 2026.04.29: InfraOcc project initiated.

📖 Abstract

Semantic occupancy prediction provides a unified 3D representation of free space, static layout, and dynamic traffic participants. Existing benchmarks and methods are predominantly designed for moving ego vehicles. Infrastructure sensors instead observe the same traffic space from fixed viewpoints, where a near-persistent static scaffold coexists with sparse and short-lived dynamic events. Treating this setting as ordinary ego-centric occupancy overlooks its distinct spatial and temporal structure.

We introduce InfraOcc, a real-world infrastructure-side semantic occupancy benchmark with dense voxel annotations in a fixed roadside coordinate system. InfraOcc provides a static-dynamic decoupled annotation pipeline and a unified protocol for camera-only, LiDAR-only, and multi-modal occupancy prediction. It also separates static and dynamic occupancy during evaluation, making the characteristic asymmetry of roadside scenes directly measurable.

Based on this observation, we propose ProSD-Occ, a progressive static-to-dynamic reasoning framework. ProSD-Occ first explains the persistent static layout, then exposes residual evidence that is not sufficiently explained by that layout, and finally recomposes static, dynamic, and free-space evidence into a unified semantic occupancy field.

✨ Highlights

• Infrastructure-side occupancy benchmark. InfraOcc reorganizes real roadside multi-modal sequences into dense semantic occupancy labels defined in a shared, fixed coordinate system.
• Static-dynamic decoupled annotation. Dynamic geometry is reconstructed from infrastructure-side LiDAR sequences and object tracklets, while the persistent background is completed using vehicle-side LiDAR observations. Vehicle-side LiDAR is used only during annotation construction.
• Unified sensing tracks. Camera-only, LiDAR-only, and camera-LiDAR models share the same voxel space, semantic labels, data split, and evaluation protocol.
• Progressive static-to-dynamic reasoning. ProSD-Occ turns the structural asymmetry of fixed-viewpoint scenes into an explicit, modality-flexible reasoning process.

🏙️ InfraOcc Benchmark

InfraOcc is built from synchronized roadside streams. The sensor platform contains four calibrated roadside cameras and two infrastructure-side LiDARs. InfraOcc contains 290 temporally continuous sequences, divided into 215 training sequences and 75 evaluation sequences.

Each keyframe is represented in a fixed roadside occupancy volume covering [-64, 64] x [-64, 64] x [-4.8, 1.6] meters at a voxel size of 0.4 meters. The resulting grid has a resolution of 320 x 320 x 16 and jointly represents semantic occupancy, free space, and unobserved regions.

The benchmark construction pipeline consists of:

1. Dynamic object reconstruction: temporally consistent tracklets and infrastructure-side LiDAR observations are aggregated in object coordinates.
2. Static background construction: moving vehicle-side LiDAR viewpoints are aligned, dynamic points are removed, and persistent infrastructure is semantically annotated.
3. Static-dynamic recomposition: static background and keyframe-specific dynamic objects are combined in the fixed roadside coordinate system.
4. Occupancy labeling: semantic voxelization and visibility reasoning produce occupied, free, and unobserved voxel states.

_{Dataset construction. Dynamic-object accumulation and static-background completion are recomposed before semantic occupancy labeling and visibility reasoning.}

📊 Static-Dynamic Asymmetry

The fixed infrastructure frame reveals two complementary patterns: static layout repeatedly occupies stable spatial regions, while dynamic participants form sparse and transient traces. The class distribution shows that this temporal asymmetry is accompanied by a strong semantic imbalance between infrastructure and traffic participants.

_{Semantic occupancy distribution. Static infrastructure dominates occupied space, while safety-critical dynamic categories form a long-tailed subset.}

🧠 ProSD-Occ

ProSD-Occ operates on a unified voxel representation produced from camera, LiDAR, or fused camera-LiDAR inputs. Its reasoning process contains four main components:

_{ProSD-Occ framework. Modality-flexible voxel features pass through static layout reasoning, residual modulation, dynamic prediction, and semantic recomposition.}

1. Static Layout Reasoner predicts a soft explanation of persistent infrastructure and estimates voxel-wise static confidence.
2. Static-guided Residual Modulation suppresses static-dominant responses while preserving complementary raw evidence through a residual path.
3. Residual Dynamic Predictor focuses on sparse, transient traffic evidence exposed by the modulated representation.
4. Semantic Recomposition adaptively combines static, dynamic, and free-space evidence into the final semantic occupancy prediction.

The modality-specific encoders and modality-agnostic ProSD-Occ reasoning head allow the same design to be used across all three InfraOcc tracks.

🎨 Qualitative Visualization

_{Camera-only qualitative comparison. Multi-view roadside images, ground-truth occupancy, representative baselines, and ProSD-Occ predictions are shown in the shared roadside frame.}

🛠️ Getting Started

1. 📦 Installation

Two tested environment guides are provided:

• CUDA 11.6 installation
• CUDA 12.8 installation

After installing the required OpenMMLab dependencies, install this repository in editable mode:

pip install -r docs/requirements/runtime.txt
pip install -r docs/requirements/tests.txt
pip install -v -e . --no-build-isolation

2. 🗂️ Dataset Preparation

Follow the dataset preparation guide to organize the InfraOcc data, generate the nuScenes-style metadata, and build multi-scale occupancy labels. The expected root directory is:

data/infraocc/
├── samples/
├── v1.0-trainval/
├── gts/
├── infraocc_infos_train.pkl
└── infraocc_infos_val.pkl

Generate the static occupancy prior required by the released ProSD-Occ configurations:

python projects/InfraOcc/preprocess/static_occ_generation.py \
  --pkl-path data/infraocc_nuscenes/infraocc_infos_train.pkl

3. 🧩 Released Configurations and Checkpoints

Input setting	Configuration	Checkpoint
Camera-only	infraocc_c_4x4_36e_prosd.py	c_prosd.pth
LiDAR-only	infraocc_l_4x4_36e_prosd.py	l_prosd.pth
Camera + LiDAR	infraocc_m_4x4_36e_prosd.py	m_prosd.pth

4. 🚀 Training

The following example trains the camera-only model with four GPUs:

bash projects/InfraOcc/configs/main_table/dist_train.sh \
  projects/InfraOcc/configs/main_table/infraocc_c_4x4_36e_prosd.py \
  4 \
  --work-dir work_dirs/infraocc_main_table/c_prosd_occ_4x4_36e

Replace the configuration and work directory with the corresponding LiDAR-only or multi-modal entry to train another track. See the main-table configuration guide for all recommended commands.

5. 🔍 Evaluation

bash tools/dist_test.sh \
  projects/InfraOcc/configs/main_table/infraocc_c_4x4_36e_prosd.py \
  projects/InfraOcc/checkpoints_update/c_prosd.pth \
  4

Use the matching configuration and checkpoint from the table above for the LiDAR-only and multi-modal tracks.

✒️ Citation

If you find InfraOcc or ProSD-Occ useful in your research, please cite:

@misc{InfraOcc,
  title={{InfraOcc: An Infrastructure Occupancy Benchmark with Static-to-Dynamic Reasoning}},
  author={Yang, Lei and Bai, Xiaokai and Li, Boqi and Lin, Chunmian and Wang, Li and Song, Ziying and Zhang, Jiahuan and Ma, Enhui and Yu, Haibao and Ma, Jiaqi and Yu, Kaicheng},
  year={2026}
}

🙏 Acknowledgement

This project builds upon the following datasets and open-source projects:

• MMDetection3D
• BEVFormer
• OpenOccupancy
• STCOcc