dronefreak/BDD100K-Weather-Classification

BDD100K: Weather Classification Dataset

Unofficial redistribution of BDD100K's per-image weather attribute, reframed as a 7-class image classification task, 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 or metadata.

This repository exists to reframe one of BDD100K's existing per-image attributes — weather condition — as a standalone image-classification task: the original detection-labelled images, grouped into class folders by the weather field already present in BDD100K's official bdd100k_labels_images_{train,val}.json. No new images, annotations, or judgments were introduced; this is a re-sort of existing, official metadata.

This redistribution is sourced directly from BDD100K's official download (registration required at https://bdd-data.berkeley.edu/), not from any third-party mirror.

---

Dataset Overview

BDD100K is a large-scale, diverse driving-video dataset. Beyond object detection, lane marking, and segmentation, every one of its 100K-image frames carries three scene-level attributes captured at collection time: weather, scene type, and time of day. This repository isolates the weather attribute into a clean 7-class image-classification dataset: clear, partly cloudy, overcast, rainy, snowy, foggy, and unknown (BDD100K's own label for frames where weather could not be determined).

Companion repositories cover the other two attributes as their own classification tasks: scene type (dronefreak/BDD100K-Scenario-Classification) and time of day (dronefreak/BDD100K-Period-Classification). All three cover the exact same 69,863 train / 10,000 validation images — only the label changes.

No test split. BDD100K's official test images (20,000) were never given attribute labels (or detection labels) at all, so they cannot be used for classification either. Only train and validation are provided here.

---

Changes from the Official Release

• Task reframed, not re-annotated. The weather value for each image comes directly from BDD100K's own bdd100k_labels_images_{train,val}.json attribute fields — the same file the detection task uses. No new labels were created; this repository only reorganizes existing ones into a classification-ready layout.
• Format converted. The official attribute is a field inside a per-split detection-label JSON. This repository instead ships one metadata.jsonl per image shard ({"file_name": ..., "label": ...}), the Hugging Face imagefolder-compatible convention, so the dataset loads directly as a labelled image-classification dataset with no custom parsing.
• Sharded for Hub limits. Images are split into shard_NNN/ subdirectories of ≤8,000 files (train: 9 shards, validation: 2 shards) to stay under Hugging Face's practical per-directory file-count ceiling. This has no effect on the data itself.
• No image pixel content was modified. No images were added or removed relative to BDD100K's official labelled set.

---

Dataset Structure

<repo>/
├── README.md
├── bdd100k_weather_banner.jpg
└── data/
    └── images/
        ├── train/
        │   └── shard_000 … shard_008/   (*.jpg + metadata.jsonl, ~8,000 per shard)
        └── valid/
            └── shard_000 … shard_001/   (*.jpg + metadata.jsonl)

where each metadata.jsonl has one line per image: {"file_name": "<name>.jpg", "label": "<weather class>"}. Hugging Face's imagefolder/load_dataset resolves file_name relative to the metadata.jsonl it's listed in and loads the image automatically as the dataset's image column, with label as a plain string column.

Splits: train 69,863 images · validation 10,000 images (79,863 total — matches BDD100K's official labelled-image counts exactly).

Classes (7)

class	train	validation	share (train)
clear	37,344	5,346	53.5%
overcast	8,770	1,239	12.6%
unknown	8,119	1,157	11.6%
snowy	5,549	769	7.9%
rainy	5,070	738	7.3%
partly cloudy	4,881	738	7.0%
foggy	130	13	0.2%

The distribution is heavily imbalanced: clear alone is over half of all images, while foggy is only 0.2% (130 training images) — too few for a reliable standalone class without resampling, augmentation, or class weighting. unknown (BDD100K's own "could not be determined" label) is a sizeable 11.6% and is a real upstream label, not a defect in this redistribution.

---

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 reorganizes an existing official attribute into a classification-ready layout; 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.