Research map / Detection and segmentation / 3D detection and driving
DiffuBox: Refining 3D Object Detection with Point Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DiffuBox, a domain-agnostic point diffusion model conditioned on LiDAR points in a normalized box view that refines noisy 3D detection boxes off-the-shelf for KITTI-to-Lyft/Ithaca365 domain adaptation.
Problem
3D object detectors degrade under domain shift (different sensors/geographies), mainly due to mislocalized boxes (wrong size, shape, orientation) rather than missed detections.
Approach
Transform points around each box into a normalized box view (scale-invariant), train a point diffusion denoiser on box-perturbed views, and refine boxes at inference by following the learned score through the chain rule, with an optional target-domain size-prior guidance term; the method is detector- and adaptation-agnostic and needs no retraining.
Key results
- On KITTI→Lyft cars with PointRCNN, Direct+DiffuBox raises 0-80m BEV mAP@0.7 from 39.06 to 59.70 and 3D mAP from 12.07 to 35.56.
- ST3D+DiffuBox reaches 66.08 BEV and 39.81 3D 0-80m mAP on KITTI→Lyft cars, improving over ST3D (55.58 BEV, 22.47 3D).
- DiffuBox improves 3D mAP for Direct across five detectors on KITTI→Lyft, e.g., CenterPoint Direct 10.29 to 34.40 0-80m.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DiffuBox box refinement to Direct PointRCNN detections | PointRCNN · KITTI→Lyft cars, 0–80m | 3D mAP@IoU 0.7: 12.07 -> 35.56 | Helped |
| Add DiffuBox box refinement to Direct PointRCNN detections | PointRCNN · KITTI→Ithaca365 cars, 0–80m | 3D mAP@IoU 0.7: 10.53 -> 20.51 | Helped |
| Add DiffuBox refinement to Rote-DA detections | PointRCNN · KITTI→Lyft cars, 0–80m | 3D mAP@IoU 0.7: 28.63 -> 41.06 | Helped |
| Add DiffuBox refinement to Rote-DA detections | PointRCNN · KITTI→Ithaca365 cars, 0–80m | BEV mAP@IoU 0.7: 42.38 -> 42.28 | Hurt |
| Use a 4x rather than 2x box-size context limit (ablation) | DiffuBox · KITTI→Lyft cars, 0–80m | BEV mAP@IoU 0.7: 55.53 -> 59.70 | Helped |
| Use a 6x rather than 4x box-size context limit (ablation) | DiffuBox · KITTI→Lyft cars, 0–80m | 3D mAP@IoU 0.7: 35.56 -> 33.61 | Hurt |
| Add shape guidance (shape weight 0.1) instead of no shape weight (ablation) | DiffuBox · KITTI→Lyft cars, 0–80m | 3D mAP@IoU 0.7: 17.54 -> 35.56 | Helped |
| Increase shape weight from 0.1 to 0.5 (ablation) | DiffuBox · KITTI→Lyft cars, 0–80m | 3D mAP@IoU 0.7: 35.56 -> 30.54 | Hurt |
Surprising findings
- Oracle localization (assigning ground-truth boxes to intersecting detections) yields large gains across adaptation methods, suggesting mislocalization dominates the domain gap.
- Retraining a detector on raw Direct predictions gives only limited gains, while retraining on DiffuBox-refined boxes gives significant improvement.
Limitations
- Does not address false negatives from completely missed detections.
- Relies on a target-domain average size for shape guidance, which requires size statistics from the target domain.
Experimental setup
Datasets: KITTI Lyft Level 5 Ithaca365 nuScenes
Benchmarks: KITTI Lyft Level 5 Ithaca365 nuScenes
Metrics: mAP@IoU 0.7 mAP@IoU 0.5 nuScenes true positive translation/scale/orientation error
Baselines: Direct Output Transformation (OT) Statistical Normalization (SN) Rote-DA ST3D PointRCNN PointPillar SECOND PV-RCNN CenterPoint DSVT
Builds on: EDM diffusion framework Output Transformation Statistical Normalization Rote-DA ST3D
Closest papers by meaning
- MonoDiff: Monocular 3D Object Detection and Pose Estimation with Diffusion Models (CVPR 2024 · similarity 0.89)
- Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.89)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.89)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.87)
- 3DiffTection: 3D Object Detection with Geometry-Aware Diffusion Features (CVPR 2024 · similarity 0.87)
- DiffLoc: Diffusion Model for Outdoor LiDAR Localization (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.