Object Detection, Loss in mAP after darknet to onnx to TRT conversion

Hardware :- Jetson Nano
DS :- 5.0
TRT Version :- 7.1.3

Hello, I have attached two files. one is results_on_darkent.csv. This files contains results on my test data when I used darkent model for inference.
another file in results_on_TRT.csv, this file contains results on the same test data when I use TRT engine for inference.

I converted the darkent to TRT this way :- darknet to onnx first and then onnx to TRT.

Question:- I’m observing precision drop for some classes after conversion. can you help me understand how/why this happens and how can this be solved ?
results_with_TRT.csv (17.2 KB)
results_with_darknet.csv (16.5 KB)

Dear @adityawasnik91,
May I know which model is used in your test.

yolov4-tiny

Dear @adityawasnik91,
Could you check if the discussion in Drop in mAP after TensorRT optimization · Issue #315 · jkjung-avt/tensorrt_demos · GitHub helps?

checking, thanks.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.