Qian Ming's Editorial ArrangementQubit report public number qbitai
The AI model is getting smaller and smaller, and the computational power is getting weaker, but the accuracy is still guaranteed.
The latest representative is an open source Chinese project just launched on GitHub:An ultra-lightweight universal face detection model.
The project contributor introduces that the size file of this model is only 1MB, and the amount of computation under 320x240 input is only 90MFlops.
Of course, the effect is not less than the current industry mainstream open source face detection algorithm, or even beyond.
The contributors to this model areLinzaiThis is a real-time, ultra-lightweight universal face detection model designed for edge computing devices or low-power devices (such as ARM reasoning):
The default FP32 precision (. pth) file size is 1.1MB, and the reasoning framework int8 is about 300KB after quantization.
The goal is to use ARM for real-time human face detection and reasoning in low-computing equipment. This also applies to the mobile-side environment (Android
According to the GitHub project page, the model has been tested in Ubuntu16.04,Ubuntu18.04,Windows 10 / Python 3.6 / Pytorch1.2 / CUDA10.0 CUDNN7.6 and other environments to ensure normal operation.
There are two versions in the model design: 1) version-slim, the trunk is slightly faster, 2) version-RFB, and the modified RFB module is added to improve the accuracy.
It also provides a pre-training model using widerface training at 320x240 and 640x480 different input resolutions, which can work better in different scenarios.
According to Linzai, there is no special operator in the whole project, which supports onnx derivation, thus facilitating the migration of reasoning.
The effect is not weaker than the current mainstream open source algorithm
What is the effect/accuracy of such a model?
Linzai also released the model's tests on accuracy, speed, scenario testing and size on the GitHub project page.
There are two contestants, one is Libface detection, OpenCV Chinese Station Master Yu Shiqi's open source face detection algorithm.
The other is Retinaface-Mobilenet-0.25 (Mxnet).
The test results on Widerface dataset are as follows:
Version-slim/version-RFB can basically achieve optimal results.
The test results in the time-consuming test of the 4-B MNN reasoning test of the raspberry are as follows:
By contrast, version-slim is not losing speed.
In the subway station, Wanda Square, office and other scenarios, version-RFB is also comparatively advantageous.
More importantly, the lightweight of the new open source model:
On the GitHub page of the project, Linzai also shared how to generate VOC format training data sets and training processes, and how to better use the model.
If you are interested, please close the portal.