Euclidean distance not a good way to compare?

[jax79sg]

Hi,

I used a small subset of LFW faces and performed a detection (crop) before feed forwarding the images into the most recent model trained with WebFaces. Next i do a pairwise Euclidean distance of each embedding and generated the attached chart. What the chart says is that both similar and dissimilar faces occupy the same distance range. How would one use distance to determine if there is a match? I also tried a KNN on the embeddings and was given a 3% accuracy. Is alignment a necessity for FaceNet?

Thank you.

[rain2008204]

@jax79sg maybe the Euclidean distance is not a good way ,so how do you solve this problem?

[LCorleone]

alignment is necessary!

[jax79sg]

Hi @LCorleone,

Thanks for the reply. I had actually deployed this on Android but has yet to find a good way to align it on Android. Would you have any recommendations?

[iamrishab]

alignment is necessary!

Even alignment doesn't help much.