I have modified the provided SSD plugin with deepstream SDK to work with caffe models. It seems to work alright but the detection performance seems a little off. It might have to do with the pre-processing operations that are applied.
Does “net-scale-factor” multiply the input image to the provided float value? for instance if a pixel value of 255 is multiplied by net-scale factor of 0.1. Will it result in a value of 25.5 for that pixel?
The plugin manual also says that the “offset” property is multiplied with the “net-scale-factor” and then subtracted from the image. I assume this means that offset values of 100;110;120 will subtract 10, 11 and 12 from every pixel in their respective channels. (Assuming net-scale-factor is set to 0.1).
Lastly, why is net-scale-factor set to a very small value for the primary detector provided with sample-app 3 inside deepstream 3.0 SDK?
Any help will be appreciated.