search

ajabri / videowalk

Repository for "Space-Time Correspondence as a Contrastive Random Walk" (NeurIPS 2020)
http://ajabri.github.io/videowalk
MIT License
267 stars 38 forks source link

Multiple cycle lengths #19

Closed vadimkantorov closed 3 years ago

vadimkantorov commented 3 years ago

@ajabri Say, we're considering a cycle I_1 -> I_t -> I_1. Do I understand correctly that the only cycle of length 1 would be I_1 -> I_2 -> I_1; and that I_2 -> I_3 -> I2 is not considered? On other words,

Thank you!

ajabri commented 3 years ago

Yes, in the current implementation, the cycles computed are I_1->I_t -> I_1 for t in [2,T]. That said, once you have computed each of the transition matrices, you can mix and match them as you like; there are certainly more learning signals that can be computed without much overhead.

vadimkantorov commented 3 years ago

Got it. Thank you!