search

MKLab-ITI / visil

Authors official PyTorch implementation of the "ViSiL: Fine-grained Spatio-Temporal Video Similarity Learning" [ICCV 2019]
Apache License 2.0
207 stars 39 forks source link

The similarity of two identical videos is not 1. #15

Closed aiminsu closed 2 years ago

aiminsu commented 2 years ago

Hey there, I was confused to find that the similarity of two identical videos was not equal to 1. Could someone help to explain it?

gkordo commented 2 years ago

Hi @aiminsu, and thank you for your question. The video similarity is calculated based on a neural network that processes two compared videos and captures the spatio-temporal relations between them to calculate a similarity score. To this end, the more spatio-temporal relations between the two videos exist, the greater the similarity score calculated by the network.

However, the network was trained to produce scores within [-1, 1] and penalized when they were out of this range. This has driven the network to generate moderate/safe estimations within the (-1, 1) range so as to not get penalized. That is why the output value is not equal to 1.

Also, I assume that by identical, you mean exactly the same (duplicate) source video file. Otherwise, any slight discrepancies between the two videos may result in a similarity drop.

gkordo commented 2 years ago

Closing due to inactivity. Feel free to re-open it for further clarifications.