What is the best way to compare vectors from imagenet? Is euclidean distance a good choice? #819
-
|
What is the best way to compare vectors from imagenet? is there a better algorithm to find similar faces? |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 1 reply
-
|
we use Euclidean distance because this is what is used by default in the open-source facial recognition libraries we use. It guarantees that the accuracy is the same as in the original library. |
Beta Was this translation helpful? Give feedback.
-
|
What are the best way to store vectors to milvus database? Does it use Euclidean distance (L2)? Or something different? How can I improve search results using milvus? Maybe somehow change vectors(normalize or something else) before inserting to milvus? What distance metric algorithm compreface uses to compare two faces? |
Beta Was this translation helpful? Give feedback.
-
|
CompreFace uses Euclidean distance (L2). Could you describe more, how bad is the milvus results? |
Beta Was this translation helpful? Give feedback.
we use Euclidean distance because this is what is used by default in the open-source facial recognition libraries we use. It guarantees that the accuracy is the same as in the original library.
There is another algorithm - Cosine similarity. Most libraries allow to use it as well. But it will make CompreFace more complex to implement and use. So we didn't add it.