I try to upgrade the tensorrt version from tensorrt3.0.4 to tensorrt4.0.6, but I find ssd detection results with int8 precision are too poor! Nothing is changed except for the *.so libs and include headers.
When I just upgrade tensorrt version from tensorrt3.04 to tensorrt184.108.40.206(RC), the results is same with tensorrt3.
I also find the calibration tables for TRT3 and TRT4 are different, 1X1 convolution layers are not included in tensorrt4.
So, is there any changes for int8 calibration/implementation in the new released TRT4.0.6??
Three int8 detection results for 3 distinct tensorrt version are here: