I have 4000 unique images of sample data from intersections and the start and end point from all lanes (approach and exit lanes) on it.
I want to train a neural network on these images to find the lanes on new images.
Someone proposed template matching via cross correlation or like segmentation via otsu and then openCV matchShape. The thing is that the arrows on the image could always look differently. Here is an image to describe what it could look like:
I have only used CNNs on the MNIST dataset and wonder whether it could be an approach for this one since one could use supervised learning.
Is it a problem to have such a detailed image? (the MNIST dataset uses just a few pixels and many other recognition mechanisms will use some kind of pooling after the segmentation which could maybe lose too many relevant details in this case?)
Also I have no idea which framework would deliver the right tools out of the box for this task