Facedetector model documentation

[google]

Facedetector model documentation

[ogl4jo3]

I'm trying to use the facedetector model (facedetector_front.tflite) directly, but there is no documentation on the output of the model. A description of the output would be greatly appreciated.

I tried to inference BlazeFace model directly, and this is python demo code. It is a simple demo. There are many shortcomings to be improved.

• The post-processing spend a lot of time, I think it was caused by python code, maybe you can use Numba to optimize it.
• And I use original NMS method instead paper's method. If you want better performance, you can implement paper's method.

[mgyong]

@yousifa We don't recommend using the facedetector_front.tflite model directly without the preprocessing and postprocessing calculators that the MediaPipe Android Face detection on GPU example demonstrates.

To understand better the details of the facedetector_front.tflite model architecture, i recommend reading the publication that the model is based on. You can also examine the code of the TfLiteTensorsToDetections calculator to better understand how to process the outputs of the facedetector_front.tflite model

Could you also provide more details on what are you trying to do with the facedetector_front.tflite model? What is the use case?

[mgyong]

Closing this issue since there hasn't been any follow up from @yousifa

[swagger2016]

I'm trying to use the facedetector model (facedetector_front.tflite) directly, but there is no documentation on the output of the model. A description of the output would be greatly appreciated.

@yousifa I font this file has been removed, can you send me a copy, my email addr is : [email protected]. many thanks

[swagger2016]

I tried the tflit file too. it only output the extracted feature. thanks for your demo. i will try it soon 发自我的 iPhone

…

在 2019年9月11日，15:38，ogl4jo3 ***@***.***> 写道： I'm trying to use the facedetector model (facedetector_front.tflite) directly, but there is no documentation on the output of the model. A description of the output would be greatly appreciated. I tried to inference BlazeFace model directly, and this is python demo code. It is a simple demo. There are many shortcomings to be improved. The post-processing spend a lot of time, I think it was caused by python code, maybe you can use Numba to optimize it. And I use original NMS method instead paper's method. If you want better performance. You can implement paper's method. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[ValYouW]

@ogl4jo3 Thank you for sharing. I did a small change to the script, instead of decoding all boxes and then filter results by score, first filter results and then decode only relevant boxes, on my laptop it reduced post-processing time from ~50ms to ~10ms
Here is the gist

[SigmaLGX]

I'm trying to use the facedetector model (facedetector_front.tflite) directly, but there is no documentation on the output of the model. A description of the output would be greatly appreciated.

I tried to inference BlazeFace model directly, and this is python demo code. It is a simple demo. There are many shortcomings to be improved.

• The post-processing spend a lot of time, I think it was caused by python code, maybe you can use Numba to optimize it.
• And I use original NMS method instead paper's method. If you want better performance, you can implement paper's method.

Thank you for sharing. I run the python demo code,but get Wrong detection box. I don't know where is the problem.