1 | initial version |
Some things which come to my mind:
The easiest way is just to sample as you like, e.g. sample only every 10th point.
As you correctly pointed out, the response won't get associated and you can't get the mass (= the response) back from the descriptor since it is typically already l2-normalized. Thus, doing this implies that you modify the code such that it doesn't get normalized then you can compute the 'strength' of a descriptor by computing its l1-norm (= mass = response).
If you don't want to modify the code you could compute the 'strength' of a keypoint afterwards, e.g. by computing the entropy of the surrounding patch.
When you want to use the dense-descriptors for classification, you can make use of local pooling, e.g. for bag-of-words you can make use of max-pooling.