Geek Answers Handbook

LSH for quickly looking for NN similarity based on distance from interference?

I am studying a quick search for NN in multidimensional vectors. (Similar to searching for similar images after selecting and calculating feature vectors)

I am currently using ORB , which describes its key points with a bit string.
A Hamming distance is required to compare two ORB descriptors.

I read taht LSH calculates its hash tables based on Eucliand Distance (L2) or Manathann distance (L1). Does this mean that LSH isn't option for comparing vectors that need Hamming distances?

Edit

LSH can work with distance from hamming because it creates a hash table based on substrings in string bit strings, therefore it works

+4
source share
2 answers

The Hamming distance is equivalent to the L1 (Manhattan) distance bounded by Boolean vectors.

+3
source

Not. LSH is not limited to L1 or L2 (Eucliand Distance). This is just an extended method of hyperplane splitting or indexing, as you call it.

LSH helps you find the most likely KNN vectors for the query vector. After that, you can use the code l1 / l2 / cosine / damming to do a detailed similarity or calculate the distance.

0
source

More articles:


All Articles