lmdb generation of Faster RCNN for INT 8

I’ve read and tested sampleINT8 in TensorRT 4. The sample is for mnist which is a classification problem. I want to apply the same procedure for Faster RCNN, but the script convert_imageset in caffe only generates db files for classification.

As a result, I want to ask:

  1. If I want to use INT 8 on Faster RCNN, are lmdb files that contain bounding boxes necessary?
  2. If bounding boxes in lmdb files are required, what is the content and format of these files?