An easy solution for face detection on CPU: how a Senior Python Engineer found an alternative to heavy AI models

For the developer, it all started with finding a tool for a specific and, at first glance, simple task.

“I was recently looking for a tool for a seemingly simple task: to understand whether there is a human face in an image,” Vladislav said on LinkedIn. “I could have taken the first popular model and closed the task. But I wanted to find a solution that wouldn’t require a GPU, a separate ML service, or a paid API.”

Additionally, it seems like it was important for the team to clearly delineate functionality:

“We needed face detection, not face recognition,” the Senior Python Engineer emphasizes. “That is, the model only answers whether there is a face and where it is located. It does not try to determine who this person is and does not store any biometric data.”

To choose the best option, Vladyslav tested several popular approaches: standard OpenCV detectors, MTCNN, YOLO, and YuNet. For an objective evaluation, he prepared a special test dataset.

“I prepared a script that analyzed the effectiveness of the data. Just before that, I created a script that downloaded 10k images for analysis, 5k with faces and 5k without,” the developer notes.

According to the benchmark results, the best performance was demonstrated by the YuNet model, a compact solution that can be easily run via OpenCV directly on the CPU.

On a sample of 10,000 images and a threshold of 0.5, the following metrics were obtained:

• Accuracy — 97.54%;

• Accuracy — 99.19%;

• Recall — 95.64%;

• F1 score — 97.40%;

• FP (False Positives) — 0.73%;

• FN (False Negatives) — 4.33%.

Thanks to this experiment, the developer received a solution that fully solves the business problem without unnecessarily complicating the architecture.

"As a result, we got a simple solution without GPUs, paid services, and unnecessary infrastructure," Vladyslav Bashtannyk concludes. "And this is exactly the case when the most powerful model is not necessarily the best. Sometimes it is much more important that the tool properly solves a specific task and does not complicate the system."

“Today, a test task can cost you a lot.” The developer urged not to launch unfamiliar repositories without analysis after discovering suspicious code

On the topic

“Today, a test task can cost you a lot.” The developer urged not to launch unfamiliar repositories without analysis after discovering suspicious code

Read the country's main IT news in our Telegram

On the topic

Read the country's main IT news in our Telegram