problems of decreasing accuracy of model quasi transformation(yolov5.onnx --> yolov5.bin )

[lilyswang]

❔Question

 Hi , @mathmanu @kumardesappan ,thanks for your nice work!   I encountered some problems when I was converting the model,(yolov5s .onnx  --->  yolov5s.bin   input_size[576x960]).  
 I use the [tidl_model_import.out(version:8.0)] to do this 。Before the quantification of the model, the position of the target box is very accurate (as shown in Figure 1 below). After the quantification, the position of the visualized target box becomes worse (as shown in Figure 2 below). I would like to ask for possible reasons. Thanks  a lot！

Additional context

[Running TIDL in PC emulation mode to collect Activations range for each layer] picture 1: [Running TIDL in PC emulation mode to collect Activations range for each layer] **[***** Calibration iteration number 0 completed ****]** picture 2 :

Convert TXT screenshot： Looking forward to your reply. ^-^

[lilyswang]

[lilyswang]

[lilyswang]

I change the numFrames to 2000 but the position of the bounding box is still poor ! (==!)