问题
In the Bag of Features/Visual Words paradigm we have a vector V in k-dimensions, where V[i]=j if the i-th centroid (obtained by k-means algorithm) is the closest one among all the k-centroids for j visual descriptors (e.g. SIFT descriptors).
AFAIK, the resulting visual vector is very sparse (it means that most of entries are 0-value) since k is really big, but my question is: what is a reasonable value for k (and so the vector size)? Hundreds of dimensions? Thousands? Especially considering that k-means execution time depends from k.
回答1:
Depends on your data, really. Here is the rule of thumb:
Too small K: your clusters will not represent for all patches. Too large K: you may get quantization artifacts and probably overfitting.
来源：https://stackoverflow.com/questions/39829213/bag-of-visual-words-what-is-a-reasonable-word-vector-dimension