Even after evaluating the network and finding the best matching parameters that can’t be cross validated using a grid search, the network still produces results that require me to move a lot of points to the correct position by hand. The image below shows the refinement GUI with the trace of points that are out of place. I am working with a large datasets of labeled images, but adding more images to the training dataset has not helped substantially.
Is it normal to have this many points of of position, and is this what I should expect for doing maDLC with eight nearly identical fish? If not, what future steps should I take to improve my model and processing?