Something to note is that since the phash algorithm is being included by default for the ignorable image directory, there is a possibility of collision as the hash is not exact.

I'm not too worried about this since it's more of a feature than a bug with perceptual hashing.

--infer-a11y-ignore-image-dir string Ignore the images in a given directory when inferring textual accessibility.

This is straightforward, nothing to say about it.

--infer-a11y-ignore-image-dir-hashes strings Hashes gathered for images in ignored images directory. (default [sha256,phash-dct])

I'm not sure if this one is clear enough. This is used to set which algorithms are used with --infer-a11y-ignore-image-dir right?

--infer-a11y-ignore-image-hashes strings Ignore the given hashes when inferring textual accessibility. Hashes are in the format :, separated by commas.

Nothing to say about this one either.

I'm interested in having @gautierchomel take on this as well.

I'm not sure if this one is clear enough. This is used to set which algorithms are used with --infer-a11y-ignore-image-dir right?

Yes

I'm not too worried about this since it's more of a feature than a bug with perceptual hashing.

OK, I just think it should be somehow made clear to users of the tool that this might cause mistakes in the evaluation

OK, I just think it should be somehow made clear to users of the tool that this might cause mistakes in the evaluation

We can cover that in the README, I've opened a separate branch for that.

--infer-a11y-ignore-image-dir-hashes strings Hashes gathered for images in ignored images directory. (default [sha256,phash-dct])

I would rename that one into --infer-a11y-ignore-image-dir-algorithms but I'm also wondering if we couldn't simply use --hash for that purpose as well.

Right now we have different algorithms used for --hash (MD5 and SHA256) and --infer-a11y-ignore-image-dir-hashes (SHA256 and pHash-DCT) but since we need to use the same algorithms to compare them, it doesn't make too much sense.

Based on our latest call with @chocolatkey:

• we'll default to SHA256 only for both --inspect-images and --infer-a11y-ignore-image-dir
• we'll use --hash as an optional flag for --infer-a11y-ignore-image-dir since there are no good reasons to have a separate flag
• if you include hashes from other algorithms with --infer-a11y-ignore-image-hashes it automatically calculates those as well, there's no need to include them with --hash as well

I'll continue working on the README branch with these new decisions in mind.

• if you include hashes from other algorithms with --infer-a11y-ignore-image-hashes it automatically calculates those as well, there's no need to include them with --hash as well

Are they then added to the Link object in the output manifest? I think this would be a weird side effect.

If I understand correctly, I feel like we're conflating the hashes that are used for processing and the ones that the user wants to output in the manifest for future consumption. Here are some ideas:

• --hash or --link-hash provides the hashes that will be added to the manifest link objects.
• We drop --infer-a11y-ignore-image-dir-hashes, instead we always use SHA256 when comparing the image-dir hashes or use whatever cryptographic hash already computed if we want to be compute-smart. But I don't think that's necessary. If there's a need to support perceptual hashes, we can do it by default or add an option --infer-a11y-ignore-image-dir-perceptual bool and use whatever perceptual algorithm we think is best in the toolkit.
• The algorithms used in --infer-a11y-ignore-image-hashes are not automatically added to the link objects in the output, unless also added to --hash.

Are they then added to the Link object in the output manifest? I think this would be a weird side effect.

No, that would require using the --inspect-images flag.

But in order to compare (and potentially ignore) images, you need to compute these hashes. That's why I feel that a single flag (--hash) makes a lot more sense.

It becomes a way to indicate what should be computed across the board.

[…] instead we always use SHA256 when comparing the image-dir hashes or use whatever cryptographic hash already computed if we want to be compute-smart.

I don't think that's a good idea because some implementers will want to use a perceptual hash in addition to a cryptographic one.

[…] If there's a need to support perceptual hashes, we can do it by default or add an option --infer-a11y-ignore-image-dir-perceptual bool and use whatever perceptual algorithm we think is best in the toolkit.

We already have two different algorithms for cryptographic hashes and another two for perceptual hashes. The list might become longer over time and implementers who have stored hashes with one of these algorithms (let's say pHash-DCT) might want to keep using it.