TartanAdversity

TartanAdversity is a synthetically augmented autonomous-driving object-detection dataset. It takes clear-weather street scenes from BDD100K and re-renders them under adverse weather (fog, rain, snow) using two different augmentation pipelines, keeping the original object annotations. It is the "fully synthetic" arm of a benchmark that studies the learnability of synthetic training data: how well detectors trained on synthetic weather generalize to real adverse-weather imagery.

Each example carries the augmented image, its provenance, and object annotations in both YOLO and COCO conventions so it can drop into either training stack.

Purpose

The dataset is designed to answer: if you augment clear driving scenes into adverse weather with a generative or procedural method, how learnable and how transferable is that synthetic data? In the accompanying benchmark, models (Faster R-CNN with ResNet/ConvNeXt backbones, and YOLO) are trained on mixtures of real BDD100K images and TartanAdversity images, then evaluated on real adverse-weather datasets (DAWN, ACDC, Foggy Zurich). TartanAdversity supplies the synthetic training material.

Dataset structure

A single train split with 36,000 images, stored as HuggingFace datasets parquet shards under data/. Each row has the following fields:

Field	Type	Description
image	image	The augmented RGB scene (resolution inherited from BDD100K, typically 1280×720).
source	string	Augmentation pipeline that produced the image — gemini or automold.
file_name	string	Original BDD100K-derived file name of the scene.
weather	string	Synthetic weather condition applied — foggy, rainy, or snowy.
yolo_annotations	struct	Boxes in YOLO convention (see below).
coco_annotations	struct	Boxes in COCO convention (see below).

The two annotation structs describe the same objects, differing only in indexing and box encoding:

• yolo_annotations — parallel lists:
  • class_id (int64): 0-indexed class, 0…9.
  • class_name (string): human-readable class.
  • bbox (float32[4]): normalized [x_center, y_center, width, height] in [0, 1].
• coco_annotations — parallel lists:
  • category_id (int64): 1-indexed class, 1…10.
  • category_name (string): human-readable class.
  • bbox (float32[4]): absolute pixels [x, y, width, height] (top-left origin).
  • area (float32): box area in pixels².
  • iscrowd (int64): COCO crowd flag (0 for all standard boxes).

Provenance axes

Every image sits at the intersection of two axes, so the set is balanced across:

• Augmentation source (source):
  • gemini — generative image-to-image re-rendering with Gemini Flash 2.0 (learned, photorealistic weather synthesis).
  • automold — procedural computer-vision augmentation via the Automold road-augmentation library (classical weather effects).
• Weather condition (weather): foggy, rainy, snowy.

This gives 2 × 3 = 6 source×weather groups spanning the 36,000 images, exactly 6,000 images per group (6,000 × 6 = 36,000).

Object categories

TartanAdversity uses the BDD100K 10-class detection taxonomy. YOLO class_id is 0-indexed; COCO category_id is 1-indexed:

YOLO class_id	COCO category_id	Name
0	1	bike
1	2	bus
2	3	car
3	4	motor
4	5	person
5	6	rider
6	7	traffic light
7	8	traffic sign
8	9	train
9	10	truck

Usage

from datasets import load_dataset

ds = load_dataset("JohnYanxinLiu/TartanAdversity", split="train")
ex = ds[0]

ex["image"]        # PIL.Image
ex["source"]       # "gemini" or "automold"
ex["weather"]      # "foggy", "rainy", or "snowy"
ex["coco_annotations"]["bbox"]        # absolute [x, y, w, h] boxes
ex["yolo_annotations"]["bbox"]        # normalized [xc, yc, w, h] boxes

# Filter to one augmentation source / weather condition:
gemini_fog = ds.filter(lambda r: r["source"] == "gemini" and r["weather"] == "foggy")

Relationship to the benchmark evaluation

In the benchmark, models are evaluated on real adverse-weather sets with a canonical weather vocabulary — clear, rainy, foggy, snowy. TartanAdversity already uses this vocabulary (foggy/rainy/snowy), so its images provide the adverse-weather training signal whose real-world transfer is being measured, directly comparable to the foggy/rainy/snowy slices reported on the real validation datasets.

Dataset creation

• Source imagery: clear-weather scenes from BDD100K.
• Annotations: object boxes are carried over from the underlying BDD100K scenes (weather augmentation does not move objects), then exported in both YOLO and COCO form.
• Augmentation:
  • Gemini — Gemini Flash 2.0 image-to-image re-rendering of the scene under the target weather condition.
  • Automold — procedural weather effects from the Automold road-augmentation library, applied with classical image operations.

Considerations & limitations

• Synthetic, not real: images are augmented approximations of adverse weather. Fidelity varies by method and condition (generative vs. procedural), and this gap is precisely what the benchmark studies — do not treat these as ground-truth real weather captures.
• Inherited BDD100K biases: geography, camera, class balance, and time-of-day distributions come from BDD100K and carry over.
• No native validation/test split: the dataset is a single train pool intended to be mixed with real data; evaluate on real adverse-weather datasets.
• Annotation origin: boxes reflect the original clear-scene labels; heavy weather effects may occasionally occlude objects that remain labeled.

Licensing

Released under CC-BY-4.0. The dataset is derived from BDD100K; users should also observe BDD100K's original terms of use and cite it. Provide attribution to TartanAdversity and BDD100K when using this data.

Citation

If you use TartanAdversity, please cite this dataset and BDD100K:

This dataset is built with Gemini Flash 2.0 (generative augmentation) and the Automold road-augmentation library (procedural augmentation); please credit both, along with BDD100K as the source imagery.

@misc{tartanadversity,
  title  = {TartanAdversity: Synthetically Augmented Adverse-Weather Driving Scenes},
  author = {Liu, John Yanxin},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/JohnYanxinLiu/TartanAdversity}}
}

@inproceedings{bdd100k,
  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 = {CVPR},
  year      = {2020}
}

@misc{automold,
  title        = {Automold: Road Augmentation Library},
  author       = {Saxena, Ujjwal},
  howpublished = {\url{https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library}}
}