Failed test on Mac: Test_Int8_nets.YOLOv4_tiny/0, where GetParam() = OCV/CPU

traffic lights in the second image may be detected

Does the CI run updated download_models.py script? Or should it be done manually? Because the model has changed and should be updated on CI.

CI failure comes from mess with .cfg file.

There is a step checking and downloading new models in download_models.py.

Looks like it cannot find the .cfg file though.

Update:
Honestly speaking, I checked the merge step of opencv_extra and it seems working well. There is an download error on the Windows host while others failed at finding .cfg file. Does it need a manually update?

I've updated the cfg file.
There is still something wrong with the int8 test, it returns different results depending quantize mode and the test can not handle it.

Batch size 1:

Batch size 1 with per-tensor quantize:

*updated images
*updated text

Lets disable int8 test with comment.

Or left it as is (restore .cfg file in opencv_extra and revert test changes).

I've fixed int8 test by slight increase of IoU match threshold and similarly for CUDA FP16 (check has not passed yet) by increasing confidence match threshold. There are sporadic problems with model download on CI.

I can see one more failure in Linux Debug build on buildbot, it can be fixed by increasing IoU difference threshold too (0.082 -> 0.118). I'll update the PR after successful CUDA test run.

Hi @fengyuentau, the download script of the Windows server always fails, I think manual upload to windows host is necessary.

Uploaded yolov4-tiny-2020-12.weights and yolov4-tiny-2020-12.cfg to the Windows host. Re-triggered the job and lets see what happens.

@zihaomu @fengyuentau @alalek Can we merge the PR?