Hacker News new | past | comments | ask | show | jobs | submit login

Google Photos has that. But I guess you mean locally?





Google photos search doesn't work well at all for me.

Even things like "[my dog name] beach" isn't reliable.

Or things like I use photos as notes. It doesn't reliably recall things like cheese when I take pictures of cheese in various stores to remember what is sold where. Not even the name of the cheese; just cheese. Ditto spices.


Don't use the name - "Mr Floppy at the beach" is meaningless, but "dog at a beach" will probably yield a lot more.

I've found google photos search to be pretty good, and if it can't find something usually the map-mode is enough to pin it down (e.g. go to the beach where you took the photo and it shows you the photos from there)


Thanks for the suggestion.

I did just check, and "dog at beach" generates sub 20% recall for me. I go to the beach weekly with my dog, take lots of photos because I'm a dork, and that first query skips many weeks.

Also, I did add my dog as a known / named pet under the explore tab, which is why I thought the name should work.

I can make it work by picking out the beach via geo, but I think the whole thing illustrates how much better this could be. I'd like to be able to get responses to queries like

* [pet name] at [beach X]

* [pet name] with sand on face

* dead seal, or even just dead animal (pics on beach)

* seaglass (recall is poor there too until I manually added to a photo album)

* dent in car

* [spice name] (I take pics of spices to know which stores offer what)

etc etc. The only way I manage the thousands of photos I have now is by carefully sorting into hundreds of albums, which google also doesn't support well.

Amongst the many many deficiencies of the app (which, tbf, does work extremely well as a read-through cache and seems to back things up very well), it likes to surface spotlights of dead people and pets. Which is not at all what I want proactively surfaced.


I think the classification engine is likely just like 500 common object classifications. It's quite old and has been around for ages so I don't think it is currently as sophisticated as modern LLMs etc. So "sand" and "face" are probably fine, but "sand on face" might be tricky. No idea though - just guessing from my own personal experience

It was also hobbled quite hard near the start as they had a scare with searches for "gorilla" accidentally returning pictures of black people so they have probably turned all the safety knobs they have up to 11, even if that impacts recall.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: