arrmlet/car-bdd-fine-grained

Fine-Grained Vehicle Detection Dataset (Corolla × BMW 3-Series, BDD100K-derived)

Real-world dashcam frames with fine-grained make annotations on top of BDD100K's existing car bounding boxes. Identifies which BDD-labeled cars are specifically Toyota Corolla sedans or BMW 3-Series sedans.

Images: 1410 unique frames · Annotations: 1524 (558 Corolla + 966 BMW 3-Series) Source imagery: BDD100K dashcam corpus (Berkeley DeepDrive)

Why this dataset

BDD100K labels every car as a single class car. We layered fine-grained make/model identification by running Qwen3-VL-8B-Instruct as a zero-shot classifier on each car crop ≥ 80×60 px, asking two yes/no questions: "Is this a Toyota Corolla sedan?" and "Is this a BMW 3 Series sedan?".

The 1,524 confident positives across 1,410 frames are this dataset.

Hit rate across 9,347 qualifying BDD car crops: ~16% (consistent with US vehicle population priors).

Schema

column	type	description
image_with_bbox	image (JPEG)	dashcam frame with bbox + make label drawn — preview only, do not train on this
image	image (JPEG, 1280×720 RGB)	clean dashcam frame — use this for training
bdd_file_name	string	original BDD filename, retained as join key
bdd_image_id	int	BDD COCO image id
bdd_ann_id	int	BDD COCO annotation id
bbox	list[float] (4)	[x, y, w, h] in pixels, top-left convention
category_id	ClassLabel	0 = toyota_corolla, 1 = bmw_3series
make	string	redundant string form for readability
weather	string	clear / rainy / snowy / overcast / partly cloudy / foggy / undefined
timeofday	string	daytime / night / dawn/dusk / undefined
scene	string	city street / highway / residential / tunnel / parking lot
width	int	image width (1280)
height	int	image height (720)

Splits

split	rows	corolla	bmw_3series	unique images
train	1,249	456	793	1,151
val	140	57	83	133
test	135	45	90	126

Splits inherit BDD100K's clip-level boundaries — no two splits share frames from the same dashcam drive. Stratification by class × weather × time-of-day is preserved.

Limitations

• Pseudo-labels from Qwen3-VL — no human verification at v0.1.0. Audit of N=50 underway; v0.2.0 will publish the noise rate.
• Class imbalance: BMW 3-Series is ~1.7× more frequent than Corolla (matches dashcam-route distribution, not corrected).
• Coverage gaps in (class × weather × tod) cells — foggy weather has exactly 1 example. Full report: https://github.com/arrmlet/car-generation/blob/main/docs/DATA_QUALITY_REPORT.md
• Geographic bias: BDD is US-only (mostly West Coast). For Ukrainian-domain coverage see the synthetic counterpart.

Companion artefacts

• Synthetic Ukrainian dataset: arrmlet/car-ukraine-synth
• Pipeline code + docs: https://github.com/arrmlet/car-generation
• Trained YOLOv11n (mAP@0.5 = 0.76 on val): see GitHub repo

Citation

@dataset{bdd_fine_grained_2026,
  author  = {Truba, Volodymyr},
  title   = {Fine-Grained Vehicle Detection — Toyota Corolla & BMW 3-Series (BDD100K-derived)},
  year    = {2026},
  url     = {https://huggingface.co/datasets/arrmlet/car-bdd-fine-grained},
  note    = {Pseudo-labels via Qwen3-VL-8B; source imagery from BDD100K (Berkeley DeepDrive).}
}

Please also cite the underlying BDD100K dataset:

@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}
}