Revision history [back]

idk, if this is a good idea (i've never seen that). the clustering is to find the most statistically relevant features, not those who "matched best" in your train data set (which might be irrelevant for a classification).
that's probably not possible or feasible. your final classification will tell, if your data processing was good or bad, and again, you cannot classify unclustered features properly (dimension problemm). also, PCA does no "comparison". i don't see, how that would be useful.
while fusing several features for machine-learning is quite common, rather "multi-voc", clustering on different datasets (using the same kind descriptor) is used. if one of your descriptor types works better than the others, stick with that. (else probably you'll degrade the ml result) also see "the curse of dimensionality"
you can use binary descriptors, like ORB, BRISK, BRIEF, etc. but you need a different algorithm for clustering them, since comparing those with L2 distance makes no sense. you'll need to find something, that uses HAMMING distance here (maybe from scikit), or write your own, like db_scan.
there have been improvements made on the BoW idea, like VLAD or fisher-vectors. look it up.

idk, if this is a good idea (i've never seen that). the clustering is to find the most statistically relevant features, not those who "matched best" in your train data set (which might be irrelevant for a classification).
that's probably not possible or feasible. your final classification will tell, if your data processing was good or bad, and again, you cannot classify unclustered features properly (dimension problemm). also, PCA does no "comparison". i don't see, how that would be useful.
while fusing several features for machine-learning is quite common, rather "multi-voc", clustering on different datasets (using the same kind descriptor) is used. if one of your descriptor types works better than the others, stick with that. (else probably you'll degrade the ml result) also see "the curse of dimensionality"
you can use binary descriptors, like ORB, BRISK, BRIEF, etc. but you need a different algorithm for clustering them, since comparing those with L2 distance makes no sense. you'll need to find something, that uses HAMMING distance here (maybe from scikit), or write your own, like db_scan.
there have been improvements made on the BoW idea, like VLAD or fisher-vectors. look it up.

it's nice, that you want to try it like this, but most likely training a cnn from scratch (or, using transfer-learning with an existing model, and your data) will be more successful.

here's a tutorial

idk, if this is a good idea (i've never seen that). the clustering is to find the most statistically relevant features, not those who "matched best" in your train data set (which might be irrelevant for a classification).
that's probably not possible or feasible. your final classification will tell, if your data processing was good or bad, and again, you cannot classify unclustered features properly (dimension ~~problemm).~~ problem). also, PCA does no "comparison". i don't see, how that would be useful.
while fusing several features for machine-learning is quite common, rather "multi-voc", clustering on different datasets (using the same kind of descriptor) is used. if one of your descriptor types works better than the others, stick with that. (else probably you'll degrade the ml result) also see "the curse of dimensionality"
you can use binary descriptors, like ORB, BRISK, BRIEF, etc. but you need a different algorithm for clustering them, since comparing those with L2 distance makes no sense. you'll need to find something, that uses HAMMING distance here (maybe from scikit), or write your own, like db_scan.
there have been improvements made on the BoW idea, like VLAD or fisher-vectors. look it up.
it's nice, that you want to try it like this, but most likely training a cnn from scratch (or, using transfer-learning with an existing model, and your data) will be more successful.