About

Wake Vision is a state-of-the-art person detection dataset specifically created for TinyML applications. It provides a comprehensive collection of high-quality images and precise annotations to train and evaluate machine learning models for efficient person detection on embedded and edge devices. Wake Vision also includes a fine-grain benchmark suite for evaluating the robustness of TinyML models.

The Dataset

Wake Vision is a large, high-quality binary image classifcation dataset for person detection:

  • Over 6 million high-quality images
  • Two training sets (Large & Quality)
  • High quality validation and test sets (~2% Label Error Rate)

Fine-Grain Benchmark Suite

Wake Vision also incorporates a comprehensive fine-grained benchmark to assess fairness and robustness across:

  • Perceived gender
  • Perceived age
  • Subject distance
  • Lighting conditions
  • Depictions (e.g., drawings, digital renderings)

Access The Dataset

HuggingFace TensorFlow Datasets Download Directly

Key Features

TinyML Focus

TinyML relevant usescase and tractable task.

Two Training Sets

One large and one high quality, ideal for data-centric AI research

Diverse Scenarios

Wide range of person detection use cases

High-Quality Test and Val

Manually labeled to ensure reliable evaluation

Model Zoo

Model Name Input Size RAM (KiB) Flash (KiB) MACs Test Accuracy
mcunet-320kb (144,144,3) 393 923.76 56,022,934 85.9%
MobileNetV2_0.25 (224,224,3) 1,244.5 410.55 36,453,732 84.7%
mcunet-5fps (80,80,3) 226.5 624.55 11,645,502 83.1%
mcunet-10fps (64,64,3) 168.5 533.84 5,998,334 82%
micronet_vww4 (128,128,1) 123.50 417.03 18,963,302 78.6%
micronet_vww3 (128,128,1) 137.50 463.73 22,690,291 78.5%
colabnas_k_8 (50,50,3) 32.5 44.56 2,135,476 77.7%
colabnas_k_4 (50,50,3) 22 18.49 688,790 75.9%
micronet_vww2 (50,50,1) 71.50 225.54 3,167,382 72.5%
colabnas_k_2 (50,50,3) 18.5 7.66 250,256 71.7%

🙋‍♂️ Contribute

Share your results with us and contribute to the leaderboard, or you can issue a PR at this link!

🏆 Challenge

The first edition of the Wake Vision Challenge has ended. Stay tuned for the next edition!

Wake Vision Challenge 🏆

Following the release of the Wake Vision Dataset, the Wake Vision Challenge was launched to advance research in TinyML. Participants were invited to contribute innovative model architectures in the model-centric track or to improve the dataset through the data-centric track. This section reports the results of the challenge. By clicking on the author name visitors can access the original submission containing the submitted model, source code, and a report describing the adopted solution. This provides a foundation for those interested in pushing the boundaries of TinyML with the Wake Vision Dataset.

Model-Centric Track

Model Name Input Size RAM (KiB) Flash (KiB) MACs Test Accuracy
ymac (Model) (50,50,3) 34.5 27.77 431,985 77.2%
samy (Model) (80,80,3) 73.5 34.55 5,718,046 79.9%
anas-benalla (Model) (50,50,3) 48.5 104.55 4,899,292 80%
mohammad_hallaq (Model) (80,80,3) 129.5 128.32 3,887,331 77.8%
apighetti (Model) (50,50,3) 48.5 278.36 693,818 74.5%
benx13 (Model) (48,48,3) 63.5 94.95 1,693,474 79.9%
cezar (Model) (50,50,3) 245.5 57.23 20,435,868 79.5%

Data-Centric Track

Author Model Name Test Accuracy
kooks wv_quality_mcunet-320kb-1mb_vww 79.2%
rgroh wv_quality_mcunet-320kb-1mb_vww 76.7%
benx13 mcunet_tiny_int8 49.6%

Example Images

Predominantly Female Person
Bright Image
Depicted Person
Young Person

License

The Wake Vision labels are derived from Open Image's annotations which are licensed by Google LLC under CC BY 4.0 license. The images are listed as having a CC BY 2.0 license. Note from Open Images: "while we tried to identify images that are licensed under a Creative Commons Attribution license, we make no representations or warranties regarding the license status of each image and you should verify the license for each image yourself."

Cite

@article{banbury2024wake,
title={Wake Vision: A Tailored Dataset and Benchmark Suite for TinyML Computer Vision Applications},
author={Banbury, Colby and Njor, Emil and Garavagno, Andrea Mattia and Stewart, Matthew and Warden, Pete and Kudlur, Manjunath and Jeffries, Nat and Fafoutis, Xenofon and Reddi, Vijay Janapa},
journal={arXiv preprint arXiv:2405.00892},
year={2024}
}

Contact

Email: emjn@dtu.dk cbanbury@g.harvard.edu AndreaMattia.Garavagno@edu.unige.it