💾 Archived View for dioskouroi.xyz › thread › 29425672 captured on 2021-12-03 at 14:04:38. Gemini links have been rewritten to link to archived content
-=-=-=-=-=-=-
________________________________________________________________________________
OP here. Here's a perceptual hash based on ML models. There's a demo at the link.
This is interesting, but I wasn't able to really exercise the demo. The first time I uploaded an image it appeared to take me to a results page but there were no matches. The subsequent two images I tested (both non-nude but racy, because I care about image classification for user-generated-content moderation) did not appear to take me to a results page at all. I'm using Firefox if it matters.
The model it uses is trained on a discrete set of categories. Somethings will not return any results, simply because there are no images indexed. Things like images of animals, dogs, a bike, a building - known objects will turn results. There is a fuller description in the blog post about what images will most likely work. There's only around half million images indexed. I'd put more but I ran out of disk space.
No matching results makes sense. That's what I got with my first attempt. But my first attempt also showed my uploaded image at the top with the caption "Uploaded Image" (and no matching images below). The subsequent couple of images I tried did not show my image with the Uploaded Image caption. Instead, after the image was uploaded I was immediately taken back to the "Find Similar Images" input page. That's why I thought there might be some sort of bug or additional input filtering.
Finally, I uploaded a randomly chosen picture of a puppy and got many other dogs and puppies. Success! :-)
I am a happy user of pHash and it was interesting to stumble across this in the submission queue. Thanks for pHash and I hope that this submission gets traction.
Good to hear. Not sure what happened on the second attempt. Sometimes it doesn't like the image and returns a 404 Error. Occasionally, it is unable to decode the image. The webapp uses tensorflow to decode the image. I've seen tensorflow randomly quit without returning while decoding images. In that event, I think it would just return to the webapp page.