arrmlet/car-bdd-fine-grained

---
license: bsd-3-clause
task_categories:
- object-detection
- image-classification
language:
- en
- uk
tags:
- vehicle-detection
- fine-grained
- automotive
- dashcam
- bdd100k
- coco
pretty_name: Fine-Grained Vehicle Detection — Toyota Corolla & BMW 3-Series (BDD100K-derived)
size_categories:
- 1K<n<10K
---

# 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](https://huggingface.co/datasets/arrmlet/car-ukraine-synth).

## Companion artefacts

- **Synthetic Ukrainian dataset:** [`arrmlet/car-ukraine-synth`](https://huggingface.co/datasets/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

```bibtex
@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:

```bibtex
@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}
}
```