Seokha Moon1,5,† · Janghyun Baek1 · Yujin Jeong2 · Daewon Chae3 · Giseop Kim4,5,‡ · Jungbeom Lee1 · Jinkyu Kim1,* · Sunwook Choi5,*
1Korea University · 2TU Darmstadt & hessian.AI · 3University of Michigan · 4DGIST · 5NAVER LABS
†Work done during an internship at NAVER LABS · ‡Work done while at NAVER LABS
* Corresponding authors
- 2026.06.18 — StreamOcc has been accepted to ECCV 2026.
- 2025.11.29 — Code released.
- 2025.11.27 — StreamOcc paper has been updated on arXiv.
StreamOcc is a real-time 3D occupancy prediction framework that streams dense voxel representations across time. It addresses two key failure modes of naive dense voxel streaming: warping distortion from temporal alignment and degraded dynamic-object representations from image-to-voxel projection.
- StreamOcc introduces a dual aggregation strategy combining StreamAgg for temporal dense voxel accumulation and QueryAgg for targeted dynamic-object refinement.
- Achieves state-of-the-art performance:
- Occ3D-nuScenes: 41.9 mIoU (+2.3 over prior SOTA / in real-time setting)
- SurroundOcc benchmark: 23.4 mIoU / 21.0 mIoUD (+1.5 / +2.0 over prior SOTA)
- RayIoU: 41.1 RayIoU (+0.8 over prior SOTA), with 34.2 / 41.9 / 47.1 at 1m / 2m / 4m
- Runs within real-time constraints (83.3 ms) and requires only 2.8 GB of GPU memory.
StreamOcc predicts voxel occupancy in a streaming manner through two complementary aggregation stages:
- Propagates dense voxel features through a recurrent streaming buffer.
- Aligns past voxel features to the current ego frame using motion-aware warping.
- Rectifies interpolation artifacts with adaptive residual refinement.
- Extracts instance-level dynamic-object semantics from image features.
- Propagates object queries over time and injects them into corresponding occupied voxel regions.
- Complements dense voxel streaming for distant, occluded, and overlapping dynamic objects.
StreamAgg and QueryAgg jointly produce a fast, memory-efficient, and high-fidelity 3D occupancy representation.
StreamOcc provides clearer and more consistent 3D occupancy predictions, significantly improving reconstruction of both dynamic objects and fine-grained static structures compared to prior methods.
StreamOcc achieves state-of-the-art performance on Occ3D-nuScenes (41.9 mIoU), the SurroundOcc benchmark (23.4 mIoU), and RayIoU (41.1), while running at 83.3 ms and using only 2.8 GB of memory.
Step 1. Set up the environment:
➡️ Install
Step 2. Prepare datasets and PKL files:
➡️ Prepare Data
# Train
bash local_train.sh StreamOcc
# Test
bash local_test.sh StreamOcc path/to/checkpointThis project is not possible without multiple great open-sourced code bases. We list some notable examples below.
If this work is helpful for your research, please consider citing the following BibTeX entry.
@article{moon2025streamocc,
title={Streaming Dense Voxel Representations for 3D Occupancy Prediction},
author={Moon, Seokha and Baek, Janghyun and Jeong, Yujin and Chae, Daewon and Kim, Giseop and Lee, Jungbeom and Kim, Jinkyu and Choi, Sunwook},
journal={arXiv preprint arXiv:2503.22087},
year={2025}
}




