README.md

January 25, 2026 · View on GitHub

DuOcc: Stream and Query-guided Feature Aggregation for Efficient and Effective 3D Occupancy Prediction

Seokha Moon1† · Janghyun Baek1 · Giseop Kim2‡ · Jinkyu Kim1 · Sunwook Choi3

1Korea University · 2DGIST · 3NAVER LABS

†Work done during internship at NAVER LABS, ‡Work done at NAVER LABS

Paper PDF

🚀 News

  • 2025.11.29 — Code released.
  • 2025.11.27 — DuOcc paper has been updated on arXiv.

✨ Highlights

  • DuOcc introduces a dual aggregation strategy combining Stream-based Voxel Feature Aggregation (StreamAgg) and Query-guided Feature Aggregation (QueryAgg) for efficient and high-fidelity 3D occupancy prediction.
  • Achieves state-of-the-art performance:
    • Occ3D-nuScenes: 41.9 mIoU (+2.3 over prior SOTA / in real-time setting)
    • SurroundOcc dataset: 23.0 mIoU (+1.1 over prior SOTA)
  • Runs within real-time constraints (83 ms) and requires only 2.8 GB of GPU memory — over 40% less memory than competing approaches.

💡 Method

Method

DuOcc is composed of two complementary components:

Stream-based Voxel Feature Aggregation (StreamAgg)

  • Aligns and aggregates voxel features over time using motion-aware warping.
  • Reduces warping artifacts via a lightweight refinement module (RefineNet).
  • Preserves spatially coherent geometry and is particularly effective for static structures, whose positions remain stable after ego-motion compensation—making them inherently suitable for stream-based accumulation.

Query-guided Feature Aggregation (QueryAgg)

  • Extracts semantics of dynamic objects from image features and encodes them into propagated instance queries.
  • Injects these instance-level query features into the corresponding voxel regions.
  • Complements fine-grained dynamic object details that are difficult to capture through voxel accumulation alone due to motion-induced misalignment, occlusion, or sparse projection.

StreamAgg and QueryAgg jointly produce a fast, memory-efficient, and high-fidelity 3D occupancy representation.

🎨 Qualitative Results

DuOcc provides clearer and more consistent 3D occupancy predictions, significantly improving reconstruction of both dynamic objects and fine-grained static structures compared to prior methods.

📊 Quantitative Results

DuOcc achieves state-of-the-art performance on Occ3D-nuScenes (41.9 mIoU) and SurroundOcc Dataset(23.0 mIoU), while running at 83 ms and using only 2.8 GB of memory, making it one of the most efficient high-performing occupancy prediction models available. These results highlight DuOcc’s strong balance of accuracy, speed, and memory efficiency, making it highly suitable for real-world autonomous driving.

🔧 Getting Started

Step 1. Set up the environment:
➡️ Install

Step 2. Prepare datasets and PKL files:
➡️ Prepare Data

🏋️ Training & Inference

# Train
bash local_train.sh DuOcc
# Test
bash local_test.sh DuOcc path/to/checkpoint

🙏 Acknowledgement

This project is not possible without multiple great open-sourced code bases. We list some notable examples below.

📃 Bibtex

If this work is helpful for your research, please consider citing the following BibTeX entry.

@article{moon2025duocc,
  title={Stream and Query-guided Feature Aggregation for Efficient and Effective 3D Occupancy Prediction},
  author={Moon, Seokha and Baek, Janghyun and Kim, Giseop and Kim, Jinkyu and Choi, Sunwook},
  journal={arXiv preprint arXiv:2503.22087},
  year={2025}
}