Fix SSIMULACRA2 weight indexing for small images #4491
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Member
|
Good catch! Probably ignoring the 1:32 scale is not the best way to handle small images; maybe it's better to relax that "min dimension 8" break condition in the loop instead (unless it's really needed because some of the operations assume such a minimum). There will still be images too small so this fix is needed in any case, but I think it would be best to keep the contribution of the low freq scales for as long as possible instead of not computing it at all, which will probably lead to scores being a bit too high for small images where that happens. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
This PR fixes what seems to be an indexing bug for small images. When calculating the average values, the loop breaks if the image has been downsampled to a point where either dimension is smaller than 8. This happens when at least one of the original dimensions is smaller than 256. If the loop breaks in this way,
Msssim::scaleshas fewer values thankNumScales. Because the channel is the outer dimension of the loop that calculates the final weighted value and the index is incremented in the loop body, different weights are used for the same scale and channel than for large images. As far as I can tell from the README of https://github.com/cloudinary/ssimulacra2, no dataset with images that small are used in the training process, so this behavior is not taken into account when tuning.Pull Request Checklist
./ci.sh lintfor automatic code formatting.Please review the full contributing guidelines for more details.