Difference in results

[Compaile]

The result of this repository and the original darknet are slightly different (a bit worse detection some objects missing).
It works 90% but isn't quite as good.

I converted the weights file with the project and loaded it.
Its all pretty straight forward, also stable but the detection is a bit worse.

I also tried to do a fixed resize before the image but that doesnt do the trick.
Any ideas where the slight difference come from?

[pfeatherstone]

What model are you using? yolov3, yolov4, ... ?

[pfeatherstone]

So dlib's upsample layers use bilinear interpolation. In darknet, the upsample layers use "nearest" interpolation. So there are at least 3 slightly different computational layers, the upsampling layers. The darknet-trained models are optimized for "nearest" interpolation, so it's expected for performance to drop slightly.

[pfeatherstone]

For yolov3 model, I think the upsampling is the only difference. For yolov4 models, i'm not sure, there could be other slight differences.

[pfeatherstone]

Do you get exactly the same performance drop if you load the model using the weights visitor directly compared to first converting the weights to dlib format, then loading that? In theory, you should get exactly the same performance drop. Otherwise, the weight converter could have an issue.

[arrufat]

I've never done a proper evaluation of the ported model. Visually, however, they seemed to work fine.
Anyway, for a few minutes now, the YOLO loss has been merged in dlib, so you can train your own YOLO detectors.
Maybe at some point, if I will try to add a nearest neighbors up-sampling layer...