dronefreak/bdd100k-period-efficientformerv2_l

EfficientFormerV2-L Finetuned on BDD100K Period (Time-of-Day) Classification

Fine-tuned EfficientFormerV2-L image classifier on the BDD100K Period (Time-of-Day) 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.

4-class time-of-day classification (daytime / night / dawn or dusk / unknown) derived from BDD100K's per-image attributes.timeofday 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-period-efficientformerv2_l", "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	81.65%
Balanced accuracy	78.36%
Macro precision	86.02%
Macro recall	78.36%
Top-1 accuracy	93.31%

Per class

Class	Precision	Recall	F1	Test images
dawn or dusk	64.21%	53.73%	58.50%	778
daytime	93.39%	95.36%	94.36%	5258
night	98.03%	98.65%	98.34%	3929
unknown	88.46%	65.71%	75.41%	35 (few)

---

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
TinyViT-21M	94.01%	83.19%	79.63%	87.95%
EfficientViT-L1	93.98%	82.68%	78.45%	88.75%
ConvNeXt-Atto	93.95%	80.75%	76.82%	86.39%
EfficientViT-B2	93.92%	82.53%	78.88%	87.49%
RepViT-M2.3	93.89%	82.35%	78.49%	87.72%
YOLO11n	93.79%	80.13%	75.19%	88.20%
EfficientViT-B1	93.77%	82.63%	78.93%	88.38%
EfficientViT-B3	93.77%	81.04%	76.29%	88.43%
YOLO11s	93.73%	81.22%	77.49%	86.42%
EfficientViT-B0	93.71%	80.98%	77.14%	86.44%
YOLO26n	93.61%	80.41%	77.72%	83.88%
MobileNetV4-Conv-Small	93.57%	80.22%	76.22%	86.06%
YOLOv8n	93.56%	80.85%	76.26%	87.87%
MobileNetV4-Conv-Large	93.46%	80.21%	74.88%	89.57%
YOLOv8s	93.46%	79.94%	75.66%	86.39%
YOLO26s	93.36%	80.10%	75.74%	86.84%
RepViT-M1.5	93.33%	81.97%	78.35%	86.95%
ResNet-18	93.32%	80.50%	76.72%	85.87%
EfficientFormerV2-L	93.31%	81.65%	78.36%	86.02%
EfficientFormerV2-S2	93.01%	80.38%	76.79%	85.35%

---

Training

Setting	Value
Epochs (max)	50
Epochs (trained)	18
Best epoch (best.pt)	8
best.pt chosen by	macro_f1 on the valid split
Early stopping patience	10
Batch size	64
Image size	224
Optimizer	auto, resolved to AdamW (peak lr 3e-04)
Weights	EMA

---

Dataset

dronefreak/BDD100K-Period-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.
• dawn or dusk is a transitional, subjectively labelled class that sits between daytime and night.

---

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{li2022rethinking,
  title={Rethinking Vision Transformers for MobileNet Size and Speed},
  author={Li, Yanyu and Hu, Ju and Wen, Yang and Evangelidis, Georgios and Salahi, Kamyar and Wang, Yanzhi and Tulyakov, Sergey and Ren, Jian},
  journal={arXiv preprint arXiv:2212.08059},
  year={2022}
}

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-period --checkpoint <weights> --data-dir <prepared dir> --split test.