Landmarks accuracy on lmk_bounds

[justusschock]

Landmarks accuracy on lmk_bounds

[justusschock]

It runs with 400 fps+ on a RTX 2080 Ti in half precision but you should be able to get significantly more fps even on your hardware. Did you try to predict or train?

I'm afraid you would have to retrain the network to your new bounds if replacing them. I had the same problem in the beginning when tracking sequences, but adding random offsets and scales to the landmarks during training und thus slightly modifying the crops did it. In this case. I might be able to retrain a network of dlbis face detector, but this should only take 3 hours.

EDIT:
Towards the Speed Issue:
I just benchmarked again with a GeForce GTX 980, 6GB also with cuda-9 (9.2), which is the most similar GPU to the GTX960M, I was able to find.I got about 120 FPS on both full and half precision.

My first random guess: Your GPU isn't fully utilized, since you have a bottleneck on your CPU providing the data.