Hacker News Reader

In a most cases, you can reshape to a 1xN^2 vector for a NxN region. This was how object detection worked long before convolutional inputs were popular.

Have a look at mnist classification using a linear SVM, for example.

View on HN

Classification makes sense, because you do a linear (or kernel) combination of the input and squash it using sigmoid to get a probability of a class. For segmentation you output a pixel mask so you would have a NxNx3 vector to predict 1 class for 1 pixel and then you would have to do it for all pixels so you'd have to encode the position as well. Alternatively, if you take a unique weight for each position you end up with get single FC layer with NxNx3 inputs and NxN outputs (N^4 parameters). I guess for me it's hard to imagine doing segmentation "back then" and I find it very fascinating.

Reply on news.ycombinator.com