Drive-SynOOD-OD

A Diffusion-Inpainted Synthetic Out-of-Distribution Object Detection Benchmark for Autonomous Driving.

One example per OOD class, inserted into real BDD100K validation scenes by diffusion inpainting, across placement positions, times of day, and weather. Red boxes are the quality-controlled ground truth. Every OOD image is paired with its unmodified base image for counterfactual comparison.

An evaluation-only benchmark for out-of-distribution (OOD) object detection in driving scenes: BDD100K validation images augmented with synthetic OOD objects (Wild Boar, Roe Deer, Other Deer, Dog, Stroller, Scooter) inserted via a diffusion inpainting model, alongside the original BDD labels and reconstructed OOD boxes. Detectors are trained on BDD100K train elsewhere and evaluated here — there is no train split.

• 24,079 images — 10,000 original .jpg + 14,079 OOD-augmented .png.
• Eval subsets: ood_positive (14,079) · ood_positive_hiconf (7,792, QA-gated high-confidence filter over ood_positive) · clean_paired (5,618) · clean_extra (4,382) · clean_all (10,000).
• OOD classes (1-based): 1 Dog · 2 Other Deer · 3 Roe Deer · 4 Scooter · 5 Stroller · 6 Wild Boar.

Contents & usage

This repository ships the benchmark as a single archive, drive-synood-od-data.tar, containing dataset/ (images, COCO/YOLO annotations, eval splits, and the FiftyOne samples.json) and provenance documents/. Download and extract it:

hf download CEAai/Drive-SynOOD-OD drive-synood-od-data.tar --repo-type dataset --local-dir .
tar xf drive-synood-od-data.tar

The current (v3) OOD annotations are also published as loose files next to the archive: ood_annotations.json and splits/ (manifest, per-subset lists, COCO views, and the ood_positive_hiconf card). They supersede the annotation files inside the archive, which date from the initial release. Pin revision v3-annotations to reproduce the paper results.

The companion code (dataloaders, detector adapters, and the evaluation pipeline) will be released on the CEA-List GitHub; the link will be added here.

License & attribution

This dataset is derived from BDD100K (© 2018 Fisher Yu / Berkeley DeepDrive), released under the BSD 3-Clause License. Please refer to the original BDD100K license terms for usage of the underlying images and labels. The synthetic OOD-augmented images and the OOD annotations added in this work are provided under the same BSD 3-Clause terms. When using this dataset, please cite both BDD100K and Drive-SynOOD-OD, and do not use the names of the original authors or UC Berkeley to endorse or promote derived work.

Citation

@inproceedings{yu2020bdd100k,
  title     = {{BDD100K}: A Diverse Driving Dataset for Heterogeneous Multitask Learning},
  author    = {Yu, Fisher and Chen, Haofeng and Wang, Xin and Xian, Wenqi and Chen, Yingying and
               Liu, Fangchen and Madhavan, Vashisht and Darrell, Trevor},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2020},
}

@inproceedings{montoya2026drivesynoodod,
  title     = {{Drive-SynOOD-OD}: A Diffusion-Inpainted Synthetic Out-of-Distribution Object
               Detection Benchmark for Autonomous Driving},
  author    = {Montoya, Daniel and Espinoza Mayzer, Mauricio Sayri and Arnez, Fabio},
  booktitle = {2026 IEEE International Conference on Vehicular Electronics and Safety (ICVES)},
  address   = {Cochabamba, Bolivia},
  month     = nov,
  year      = {2026},
  publisher = {IEEE},
}