no, it’s not meant to work that way. you should collect a dictionary of statistically relevant features from a ton of unrelated images, an offline task.
then you match your image features against that dictionary (instead of the other image), in the hope, to retrieve more significant results
no way. not even with correct DescriptorMatchers.
kmeans clustering is using L2 distance, thus you cannot use binary / bitstring features.
far too less, to be useful
(~2000, maybe, and yes. clustering will take decades …)
remember, that this is the final desc size for the similarity comparison
use an empty Mat to push_back() to, else you have an uninitialized 1st row !
are you sure ? and line 139, is in your prog, no ?
i’d rather think, it’s from the next line, bowDE.compute() , where you got the signature totally wrong, look:
(should be something like:
Mat finaldesc;
bowDE.compute(descriptors1, finaldesc);
)