DeepStream return not same mask as TensorRT int8 engine

[rashit.bayazitov.1995]

Please provide complete information as applicable to your setup.

• Hardware Platform GPU
• DeepStream Version 5.1-21.02
• TensorRT Version 20.12
• NVIDIA GPU Driver Version 450.142.00
• Issue Type bugs
• Reproducing, create HRNet, convert it in int8 TensorRT, then run it on examples and compare the tensorRT and DeepStream mask

So i convert my model in int8 with TensorRT and compare masks from TensorRT engine runner with DeepStream, and notice the masks not the same.

On left picture (int8 TensorRT inference), you can see in the middle of image small yellow blob inside big blue blob (approximate coordinates for small blob x = 650, y = 225), that blob is very important in my task. On the over hand on right picture that blob doesn’t exists.
The question is, how can i get the same mask as in TensorRT?

[mchi]

HI @rashit.bayazitov.1995 ,
Sorry for delay!
I think, we may need to dump their TensorRT input data to do comparision to check if it’s caused by different pre-processing.

So, could you refer to 2. [DS5.0GA_Jetson_dGPU_Plugin] Dump the Inference Input section to dump the DeepStream TRT input and campare with the TRT input of the standalone sample ?