dronefreak/bdd100k-scenario-resnet18

ResNet-18 Finetuned on BDD100K Scenario Classification

Fine-tuned ResNet-18 image classifier on the BDD100K Scenario Classification dataset, trained and evaluated as part of BDD100K-Toolkit, a dependency-clean toolkit for preparing BDD100K, training models on it and evaluating them with the same metrics on the same splits.

7-class driving-scenario classification (city street / highway / residential / parking lot / gas stations / tunnel / unknown) derived from BDD100K's per-image attributes.scene field. Unofficial task; follows the Kaggle dataset of the same name.

---

Usage

import timm, torch
from huggingface_hub import hf_hub_download
from PIL import Image
from torchvision import transforms as T

ckpt = torch.load(
    hf_hub_download("dronefreak/bdd100k-scenario-resnet18", "best.pt"), map_location="cpu", weights_only=True
)
model = timm.create_model(
    ckpt["model_name"], pretrained=False, num_classes=len(ckpt["class_names"])
)
model.load_state_dict(ckpt["state_dict"])
model.eval()
prep = T.Compose(
    [T.Resize((ckpt["imgsz"],) * 2), T.ToTensor(), T.Normalize(ckpt["mean"], ckpt["std"])]
)
with torch.no_grad():
    probs = model(prep(Image.open("street.jpg").convert("RGB"))[None]).softmax(1)[0]
print(ckpt["class_names"][probs.argmax()], f"{probs.max():.1%}")

---

Results

Evaluated on the test split (10000 images).

Metric	Value
Macro F1	47.29%
Balanced accuracy	44.83%
Macro precision	54.25%
Macro recall	44.83%
Top-1 accuracy	77.14%
Top-5 accuracy	99.91%

Per class

Class	Precision	Recall	F1	Test images
city street	79.69%	86.67%	83.03%	6112
gas stations	0.00%	0.00%	0.00%	7 (few)
highway	74.39%	69.39%	71.80%	2499
parking lot	52.17%	24.49%	33.33%	49 (few)
residential	67.15%	51.72%	58.43%	1253
tunnel	77.78%	77.78%	77.78%	27 (few)
unknown	28.57%	3.77%	6.67%	53

This checkpoint classifies none of the test images of gas stations (7 images) correctly, so treat it as unsupported.

---

Model Zoo

All runs below were evaluated on the same test split, sorted by top-1 accuracy.

Model	Top-1	Macro F1	Balanced acc	Macro precision
EfficientViT-B3	79.32%	57.70%	53.59%	71.54%
YOLO11n	78.58%	49.47%	46.05%	60.98%
YOLO11s	78.58%	49.18%	45.24%	61.98%
MobileNetV4-Conv-Small	78.20%	52.89%	48.69%	61.41%
MobileNetV4-Conv-Large	78.15%	50.00%	45.48%	63.03%
EfficientViT-B1	78.01%	54.15%	49.41%	70.13%
TinyViT-21M	77.99%	62.52%	66.21%	59.66%
YOLO26s	77.81%	52.24%	49.23%	58.28%
ConvNeXt-Atto	77.44%	61.06%	60.34%	67.92%
YOLOv8s	77.15%	48.43%	46.64%	51.16%
ResNet-18	77.14%	47.29%	44.83%	54.25%
YOLO26n	76.78%	48.33%	43.94%	58.99%
ConvNeXt-Tiny	76.69%	56.88%	59.82%	54.60%
YOLOv8n	76.13%	46.40%	43.26%	54.14%
EfficientViT-B0	75.94%	46.10%	42.64%	53.74%

---

Training

Setting	Value
Epochs (max)	50
Epochs (trained)	24
Best epoch (best.pt)	14
best.pt chosen by	macro_f1 on the valid split
Early stopping patience	10
Batch size	128
Image size	224
Optimizer	auto, resolved to AdamW (peak lr 3e-04)
Weights	EMA

---

Dataset

dronefreak/BDD100K-Scenario-Classification holds the prepared splits these models were trained and evaluated on.

---

Limitations

• Unofficial task: labels are BDD100K's per-image attributes, not a benchmark with a public leaderboard, so scores are only comparable with other models evaluated on this split.
• Not the official test set: test here is BDD100K's official validation split (the official test labels are not released) and valid is a seeded 15% slice of the official train split.
• Heavily imbalanced: rare classes have very few test images, so their per-class scores are noisy. Prefer macro F1 over accuracy.
• Images are US dashcam frames; generalization to other regions or camera setups is untested.
• BDD100K is released for non-commercial research and education. Check its license before any use of these weights beyond research.
• scene is a coarse per-image label; frames showing several settings (e.g. a highway entering a tunnel) get a single class.

---

License

The weights are released under the license in the metadata above. They were trained on BDD100K, which is free for non-commercial research and education; commercial use needs separate permission (see https://www.bdd100k.com/). The prepared dataset on the Hub is tagged license: other, and the original BDD100K terms still apply.

---

Citation

Model

@article{he2015deep,
  title={Deep Residual Learning for Image Recognition},
  author={He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian},
  journal={arXiv preprint arXiv:1512.03385},
  year={2015}
}

Dataset

@article{yu2018bdd100k,
  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},
  journal={arXiv preprint arXiv:1805.04687},
  year={2018}
}

---

Files

• best.pt
• metrics.json
• results.csv
• assets/demo_banner.mp4
• assets/demo_banner_poster.jpg
• README.md

---

Reproduce

Trained and evaluated with BDD100K-Toolkit: bdd100k-evaluate --dataset bdd100k-scenario --checkpoint <weights> --data-dir <prepared dir> --split test.