How to compare 2 ArcFace features?

[pango99]

I use cosine similarity to calculate 2 features similarity,but I am not sure whether it is the best way, so whether cosine similarity is suite for arcface's feature comparing? Is there any other ways to do the feature comparing?

[cyrojyro]

you can use sum of squared error
np.sum(np.square(f1 - f2))
(from deploy/test.py)

[pango99]

you can use sum of squared error
np.sum(np.square(f1 - f2))
(from deploy/test.py)

thank you, how accurate is it ? the formuler looks like euler distance

[Walstruzz]

U can also use Cosine Similarity.

# if embedding_1.shape is (1, emb_size,), use:
embedding_1, embedding_2 = normalize(embedding_1), normalize(embedding_2)
sim = embedding_1 @ embedding_2.T

-1 <= sim <= 1.

[pango99]

U can also use Cosine Similarity.

# if embedding_1.shape is (1, emb_size,), use:
embedding_1, embedding_2 = normalize(embedding_1), normalize(embedding_2)
sim = embedding_1 @ embedding_2.T

-1 <= sim <= 1.

thank you,but I am a c language programmer,I'm using mxnet c api to use ArcFace model,and I'm a python language starter,so I cant understand the meaning of below line code:
sim = embedding_1 @ embedding_2.T

what's the meaning of "@" symbol?

[Walstruzz]

Sorry, @ equals to dot in numpy.
For mxnet NDArray, use nd.dot(embedding_1, embedding_2.T). :-)

[yuanbit]

@pango99: how do you extract ArcFace feature vectors using your own dataset and the pretrained models?

[pango99]

@pango99: how do you extract ArcFace feature vectors using your own dataset and the pretrained models?

use MXNet C API

[lebionick]

@pango99
same question, it's not obvious from the formula of loss, but these guys just use cosine dist
https://www.groundai.com/project/arcface-additive-angular-margin-loss-for-deep-face-recognition/1

During testing, the score is computed by the Cosine Distance of two feature vectors. Nearest neighbour and threshold comparison are used for face identification and verification tasks.