Are the values extracted by SURF descriptor a good input for neural network?

Question

As mentioned in the title, are those values good enough to train a neural network for classification tasks (2 classes)?

My objective is to try segmenting foreground and background using feature points extracted using SURF. These feature points extracted will be input into a neural network (supervised or classification).

My question is, are these values (128 or 64 gradient information) a good choice to train a NN?

Answer 1

Short answer: Yes, very good.

Slightly longer answer: You're relying on SURF to do what SURF does best and you're relying on the NN to classify/recognize the pattern of those inputs. That's exactly right. In reality, it always depends on the exact task at hand and if, for whatever reason, SURF doesn't pick up the features relating to the model you're building, then you'll have trouble. But in general: thumb's up.