Input image resolution for OpenVino toolkit

[Nathwani__Abhijit]

Hi Everyone,

With some tinkering around with the Intel OpenVino toolkit, I was thinking about what should be the ideal resolution of the video feed to the sample programs, so that the output is optimal? I have a camera with low resolution. Will that work fine?

I'm new in this domain and hence please don't mind the question,

Thanks,

Abhijit Nathwani

[Shubha_R_Intel]

Dearest Nathwani, Abhijit,

The answer to your question is - if it doesn't bother Tensorflow/Onnx/MxNet/Caffe/etc...then it won't bother OpenVino for FP32. OpenVino's Model Optimizer loses negligible accuracy during model conversion to FP32. However, for FP16 (used by GPU and Movidius devices), some accuracy may be lost but there's no way to guesstimate unless you build an inference application and see your results. 

You can also perform inference at lower precision using Tensorflow, and see how that works. If it works at FP16 on Tensorflow, then it should work fine on OpenVino.

The message I'm hoping to convey is that OpenVino doesn't care about input resolution. Low input resolution matters more during the training phase. But if your model does a good job in Tensorflow (recognizing low-res images or video) then it will also work just fine in OpenVino.

Hope I answered your question,

Shubha