search

terryky / tflite_gles_app

GPU accelerated deep learning inference applications for RaspberryPi / JetsonNano / Linux PC using TensorflowLite GPUDelegate / TensorRT
MIT License
488 stars 130 forks source link

BlazePose doesn't work with quantized Int8 model #33

Open enricoantonini84 opened 1 year ago

enricoantonini84 commented 1 year ago

Hi, I've tried to run, on Debian 11 following the installation instruction, the script that performs pose detection w BlazePose on the included images and it works fine. However, if I try to run the script with -q parameters to load the Int8 model, no pose is detected on any images. Is there anything else I should do to make the detection work with the quantized model?

Regards Enrico

terryky commented 1 year ago

Yes, you are right.

Unfortunately, some quantized models included in this repository may be low accuracy. I think there are several techniques to improve accuracy, but I am not familiar with them.