BDD100K 2018: Object Detection Dataset

Unofficial redistribution of BDD100K's 10-class object-detection task, reformatted into a standardized YOLO-compatible layout, under BDD100K's own data license (non-commercial/educational/research redistribution, with notice, explicitly permitted).

Disclaimer

This repository is not an official release of BDD100K.

BDD100K was created by Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell at UC Berkeley (BAIR), who retain all copyright. This repository does not claim ownership of any images, annotations, or metadata.

Two-hop provenance. This redistribution's raw files came from the Kaggle mirror solesensei/solesensei_bdd100k (uploaded by Kaggle user SoleSensei, not by Berkeley DeepDrive). That mirror declares no license of its own for the data (Kaggle lists it as "Other (specified in description)"), so it changes nothing about the terms that actually govern this data — see License below for the real, verified terms and why they permit this mirror.

Companion repositories cover BDD100K's per-image scene attributes as separate image-classification tasks: weather, scenario, and period/time-of-day.

Label vintage: 2018, not the CVPR 2020 paper year. The BDD100K paper was published at CVPR 2020, but the detection labels redistributed here are BDD100K's original 2018 annotation release (matching the copyright year on the data license itself, reproduced in full under License) — carried through the Kaggle two-hop mirror described above. We looked for a separate, later-dated ("2020") label release to confirm whether the annotations had been revised since 2018, and could not locate or verify one; if BDD100K ever shipped a distinct 2020 label set, it is not what this repository contains. Treat the labels here as the 2018 vintage, independent of the paper's 2020 publication date.

---

Dataset Overview

BDD100K is a large-scale, diverse driving dataset spanning 10 tasks; this repository covers its 100K-image object-detection task specifically: 79,863 labelled 1280x720 dashcam images (69,863 officially labelled train + 10,000 officially labelled val; the official held-out test split of 20,000 images has no released labels at all and is not used here) annotated across 10 classes spanning vulnerable road users (person, rider), vehicles (car, truck, bus, train, motor, bike), and traffic control (traffic light, traffic sign), captured across diverse weather, time-of-day, and scene conditions in the US.

---

Changes from the Official Release

• Official val kept as test. BDD100K's official labels are only released for train (69,863 images) and val (10,000 images) — the real test split is unlabelled. This repository uses the official val split, unchanged, as the canonical test split here (consistent with it only being usable as a held-out eval set, not for further training).
• A validation split was carved out of train. Since there is no official train/validation split, this repository holds out a seeded random 15% of the official train images (seed 42) as valid, leaving 59,384 for train. This is a random image-level split; BDD100K images are single dashcam frames (not continuous video sequences within this release), so unlike some driving datasets this does not risk near-duplicate consecutive frames crossing the split boundary.
• Format converted. The official ground truth is a per-split JSON with one dict per image mixing box2d detection boxes with poly2d lane-marking/drivable-area annotations (a different task). Only box2d entries are converted here; poly2d entries are dropped. Boxes are exported both as YOLO-normalized .txt labels and as absolute-pixel COCO [x, y, w, h] boxes in a metadata.jsonl per image shard.
• Sharded for Hub limits. Images are split into shard_NNN/ subdirectories of ≤8,000 files to stay under Hugging Face's practical per-directory file-count ceiling (train: 8 shards, valid: 2, test: 2). This has no effect on the data itself.
• No image pixel content was modified. No boxes were resized, moved, or reinterpreted.

---

Dataset Structure

<repo>/
├── README.md
├── bdd100k_banner.jpg
└── data/
    ├── data.yaml
    ├── images/
    │   ├── train/   shard_000 … shard_007/   (*.jpg + metadata.jsonl, ~8,000 per shard)
    │   ├── valid/   shard_000 … shard_001/
    │   └── test/    shard_000 … shard_001/
    └── labels/
        ├── train/   shard_000 … shard_007/   (*.txt, mirrors images)
        ├── valid/   shard_000 … shard_001/
        └── test/    shard_000 … shard_001/

where:

• data/images/<split>/shard_NNN/ contains the 1280x720 RGB dashcam images, plus a metadata.jsonl (file_name and objects.bbox as absolute-pixel COCO [x, y, w, h] with objects.categories) that drives the Hugging Face dataset viewer.
• data/labels/<split>/shard_NNN/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized), mirroring the image shard layout.
• data/data.yaml is the Ultralytics dataset configuration file (class names, split paths, relative to data/). Ultralytics resolves the shards automatically.
• Splits: train 59,384 images / 1,091,589 boxes · valid 10,479 images / 195,282 boxes · test 10,000 images / 185,526 boxes (79,863 images / 1,472,397 boxes total — matches BDD100K's official labelled-image counts exactly).

Classes (10)

id	class name	instances (all splits)	share
0	person	104,611	7.1%
1	rider	5,166	0.4%
2	car	815,717	55.4%
3	truck	34,216	2.3%
4	bus	13,269	0.9%
5	train	151	0.0%
6	motor	3,454	0.2%
7	bike	8,217	0.6%
8	traffic light	213,002	14.5%
9	traffic sign	274,594	18.6%

car dominates at over half of all boxes; train is vanishingly rare (151 instances total). See docs/datasets/bdd100k in the DetectionBench repository for box-geometry statistics.

---

Dataset Sources

Original Paper

BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning

Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, Trevor Darrell

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pages 2633-2642. DOI: 10.1109/CVPR42600.2020.00271

Official Resources

• Official Website: https://www.vis.xyz/bdd100k/
• Official Download Portal: https://bdd-data.berkeley.edu/
• Official Code Repository: https://github.com/bdd100k/bdd100k

---

Attribution

All credit for collecting and annotating this dataset belongs entirely to the original BDD100K authors: Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell, and UC Berkeley / BAIR.

This repository only converts their box2d detection annotations to a YOLO-compatible layout and adds a validation split; it does not modify, reinterpret, or take credit for the underlying imagery or labels.

If you use this dataset in your research, please cite the original publication below.

---

License

BDD100K's code repository is BSD-3-Clause, but that license applies only to the code, not the data. The data and labels (downloaded from https://bdd-data.berkeley.edu/) carry their own license, reproduced here in full as required by its own terms:

Copyright ©2018. The Regents of the University of California (Regents). All Rights Reserved.

THIS SOFTWARE AND/OR DATA WAS DEPOSITED IN THE BAIR OPEN RESEARCH COMMONS REPOSITORY ON 1/1/2021

Permission to use, copy, modify, and distribute this software and its documentation for educational, research, and not-for-profit purposes, without fee and without a signed licensing agreement; and permission to use, copy, modify and distribute this software for commercial purposes (such rights not subject to transfer) to BDD and BAIR Commons members and their affiliates, is hereby granted, provided that the above copyright notice, this paragraph and the following two paragraphs appear in all copies, modifications, and distributions. Contact The Office of Technology Licensing, UC Berkeley, 2150 Shattuck Avenue, Suite 510, Berkeley, CA 94720-1620, (510) 643-7201, otl@berkeley.edu, http://ipira.berkeley.edu/industry-info for commercial licensing opportunities.

IN NO EVENT SHALL REGENTS BE LIABLE TO ANY PARTY FOR DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, INCLUDING LOST PROFITS, ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN IF REGENTS HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

REGENTS SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE SOFTWARE AND ACCOMPANYING DOCUMENTATION, IF ANY, PROVIDED HEREUNDER IS PROVIDED "AS IS". REGENTS HAS NO OBLIGATION TO PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.

Source: https://github.com/bdd100k/bdd100k/blob/master/doc/source/license.rst

Accordingly:

• Redistribution for educational, research, and not-for-profit purposes is explicitly permitted, without fee, provided this notice is carried forward — which is what this repository does.
• Commercial use and distribution is restricted to BDD and BAIR Commons members and their affiliates. Contact UC Berkeley's Office of Technology Licensing (contact details above) for commercial licensing.
• This repository, and any further redistribution of it, must carry this same notice.

---

Citation

If you use this dataset, please cite:

@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={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2633--2642},
  year={2020}
}

---

Acknowledgements

We sincerely thank Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell, and UC Berkeley / BAIR, for creating and publicly releasing this valuable driving-scene benchmark.