I am trying to predict the position (x,y)-coordinates of 16 peaks in an image.
The images look like the following:
The green dots are the true peak locations and the red crosses represent a prediction of this network:
I have tried out many configurations of differnt Kernel sizes, number of layers, less/ more dense/ convolutional layers and so on. I use Adam (lr=e-3) and a dropout of 20% here but I also have tried different configurations in this parameters. I train on a batch size of 64 and aim to regress the relative coordinates (also tried with absolute coordinates).
It seems that the network is not able to train the ability of recognising a peak; instead it regresses always a kind of average.
At the moment I lack of further ideas. So I hope someone can give me a hint of what I am doing so wrong. Can you imagine of a reason for my problems?
