I am almost at the end of week 4 of the Convolutional Neural Networks module, in the Deep Learning course. After seeing a post from Yann Lecun on image segementation, I got to thinking. If you were to produce lots and lots of training data where sailboat images were labelled like, “mast”, “rudder”, “jib”, “boom”, could you train a neural network to look at some arbitrary sailboat and segment it into component parts? Seems like you could. Like segmenting a face into eyes, ears, nose, etc. How much effort would this take? Sorry for the naive question, but, after all, I am only a “freshman” in Coursera U.
My initial reaction is this falls into the ‘just because you can, doesn’t mean you should’ category. What would you do with it? My intuition is that it would be a lot harder to segment parts of the boat than to segment an image into sailboat, water, buoy, dock etc. because of the number of parts with challenging shapes - boom, cleat, halyard - ugh.
Regarding how much data is required to train a CNN, there have been a few threads on the topic in this forum. Here is one of mine:
2 Likes