Comment by benbristow
5 years ago
Your homepage says "protect your photos/faces etc. from algorithms"
The algorithms are what makes Google Photos; Google Photos. If I wanted to just store my photos I'd throw them in a S3 bucket or Dropbox or something.
Google Photos lets me automatically categorise my photos by person, lets me search my library using text search for anything (e.g. I can search 'museum' and see pictures I've taken in museums). That is where the real value of Google Photos comes into play.
> But we are far from where we want to be in terms of features (object and face detection, location clustering, image filters, ...) and user experience. We are hoping to use this post as an opportunity to collect feedback from fellow hackers.
So you're going to implement algorithms then?
> So you're going to implement algorithms then?
Yes, we will implement the algorithms, purely on the client side, such that we don't hold indexes to your personal data.
But I understand how that piece of text could have thrown you off, I'll think of ways to rephrase it. Thanks for pointing it out.
Actually I'm really curious how you do this. If the photos aren't stored client side, then how do you search? Do you have a thumbnail of every photo client side? Is that enough? I mean ImageNet scores are still pretty low for small/fast neural nets. And ImageNet isn't even representative of real world photos. So obviously to be successful you're going to have to continue training. So how do you do this in a privacy preserving way? Even federated learning can have some issues because images can be reconstructed from gradients.
> Do you have a thumbnail of every photo client side
In the happy path the files/thumbnails are indexed before they are uploaded. But we are designing a framework that will pull files/thumbnails for indexing if they are unindexed or indexed by older models.
> how do you do this in a privacy preserving way
Our accuracy will not match that offered by services who index your data on their servers. But there's a trade off between user experience and privacy here, and we are hopeful that ente will be a viable option for an audience who is willing to sacrifice a bit of one for a lot of the other.
22 replies →
You can run algorithms locally and still violate privacy by uploading private facts derived from the data with algorithms. Saying you won’t hold “indexes” doesn’t begin to cover it.
Well, it does begin to cover it. Do you have to be so strident?
1 reply →
But that will mean that for every version of the algorithms, it have to read all the photos since 15 years ago... my phone battery will die soon.
And if I need to have other kind of client... like a nas to do that... Why I need the cloud?
> phone battery will die soon
Indexing will be opt-in. You will be able to run the indexing only on your desktop client for instance.
> Why I need the cloud?
So that you don't have to manage your own storage infrastructure? But if you would like to do that, then there are self-hosted alternatives that will better serve your use case.
Agree with the above poster. I don't care about algorithms. I want algorithms. But I want algorithms that only work for me. Screw off everyone else.
Apple used to sell this. Then they stopped.
Those "algorithms" can run locally, on a NAS or a desktop, generate the metadata and make it available to you only on your mobile.
I can see myself paying for such software if it was mature enough.
Synology Photos is one such solution already for example.
I have Synology, actually. Is Synology Photos trustable?
9 replies →
> Those "algorithms" can run locally
But I don't want my GPU burning away running them when they could run much more efficiently and out of mind in the cloud.
Then you aren't the target audience?
Am I the only one who never realized you can search "museum" and see your museum photos?
Now that you've mentioned it, yes, I'd like to try that. But as a counterpoint to your argument, I've never needed it, and I suspect that a lot of people may not actually be getting the same value propositions that you're getting.
On the other hand, Google Photos is Google Photos. But it's often a mistake to compete directly with an established product. New ideas tend to win by transcending the competition.
I propose that if this Show HN turns into a product, it will be because it does something people didn't realize they wanted. Maybe that's privacy. I don't know.
I use it all the time - it's the killer feature of google photos. The premise is that if you come back from vacation with 300 photos, it's unlikely that you (the average non photography-nerd user) are going to sit there and tag them all. If in a few years you want to find "that photo of me you took on the beach in north carolina", with a quick search you can.
There are annoying limitations though, probably because the original team moved on and it's in maintenance stage. Using my example above, google photos has no idea what the "outer banks" are (which is where the beach photos were taken in north carolina) and returns no results. It also has trouble parsing out entities from search terms, so "north carolina beach maggie" isn't going to find pictures of Maggie on the beach in North Carolina (which you'd think they could really fix given that, well, they're google). Finally, there's no way (that I know of) to jump from search results to your full timeline; let's say that "north carolina beach" gets me a bunch of beach pictures from January 2015 (yeah, it was cold), but doesn't have _the_ picture from the trip that I know I want - there's no direct way to click to January 2015 from the results, which really sucks. (Instead you have to go back out of results and use their fiddly scroll to get there.)
Yeah, it's a killer feature, but I really wish they had some sort of a documented "search API".
Instead of natural language search, where I have no idea whether it understood me, I wish I could do (modifying your example):
"North Carolina" "Maggie Thomson" "Tom Morgan" -beach 2018
for all photos in NC, with Maggie and Tom, not in a beach from 2018
and even better, if it could tell me the number of results that would show up if we removed each keyword above.
I guess it's a tough problem, even for Google :(
> there's no direct way to click to January 2015 from the results, which really sucks. (Instead you have to go back out of results and use their fiddly scroll to get there.)
It's amusing how people's insights can turn myopic. Search in photos is the killer feature, and it even solves the problem that you have.
If you realize that you need to see photos from January 2015, don't try to scroll back in your photos feed. Just do a second search for "January 2015".
2 replies →
I try to use it often but it works pretty poorly and I always have to scroll through years of photos to look for what I need.
For me the killer feature of Google photos are: - Free storage of photos (hence why I'll move after I run out of free space) - Tagging faces - Sharing albums
It's a great idea that works in a limited way. Getting that next 30% is going to take awhile nevermind natural language queries.
Nah, a couple gigs of free storage is Google's killer feature. Photo organization is braindead simple and barely requires more than S3.
There's more you can do honestly. Search and and assign people so you can find picture with just them. This also works for pets. People, pets , objects, place, etc. Hell, I searched the car I use to drift and it showed up. It's really neat.
The search is really quite fun to play with, and very useful! I also like searching on the map and seeing where I’ve taken photos. Especially if I’m looking for one particular photo, it’s fun to zoom in from the world map
Thanks for pointing that out. I actually had the opportunity to sync my iPhone photos to Google Photos, but opted to decline. This made me reconsider; cheers.
4 replies →
I use this feature occasionally, but it also seems to be pretty bad for the searches I try. For example, if I search for 'dog', I do indeed get pictures back that contain my dog. However, there are a ton of false negatives -- that is to say, the 'dog' search doesn't show me all of the photos that most definitely and very clearly have my dog in them.
And it's not just dogs. Specific people, locations (before I turned of geotagging on my photos), scenery (mountains, outdoors), etc.
Sometimes this search is nice, but it's not good enough that I can really rely on it.
We need to make this stuff local again, that will be the real competitor to big corp Foo... no servers, no end-to-end, no service cost, no ads, no privacy issues, no random revokation of accounts without recourse, just one end - the users. We can have face detection etc locally if people want it... cycles, it's going to happen eventually.
we had that, but almost everyone decided they like the cloud better.
>lets me search my library using text search for anything
This is untrue, and actually one of the reasons I hope a strong competitor to Google Photos comes along soon. The search function is, for whatever reason, heavily censored and perhaps even biased in some circumstances. Worse, it is completely useless. For example, the query "fat" returns nothing, despite the fact that my gallery is filled with drawing reference photos that includes plus-sized people. "Black people" returns photos of non-black people, and (infamously, and perhaps for related reasons re: the shortcomings of Google's image recognition and tagging algorithm) "gorilla" returns 0 results. "Red shirt" returns an image of a blue decorative screen; "comic" returns anime and webpage screenshots; "woman" returns multiple photos consisting entirely of groups of men.
The situation is dire.
Think of it from Google's POV. Imagine if the tabloids found out about a situation of someone searching for 'fat' in the search bar and then it coming back with pictures of themselves or their friends - that could cause some serious controversy.
Well, this gets to the heart of one of the issues with the current approach to AI. Statistical consensus doesn't always align with a user's personal view or desires. I don't know how you solve the problem; my issue is that Google doesn't seem to know, either, but they insist that they do.
To think that someone can just throw their photos in s3 assumes people are ops, devils, or devs. That’s a small slice of the population. What about everyone else?
I also mention Dropbox. I haven't used it for a while though
You're right, a few hours of work on top of S3 are needed to obviate Google Photos.
Besides search another feature of Google Photos that I would need is automatically inclusion of photos in shared albums based upon who is in them. Some examples:
I have an album shared with my parents which photos of my daughter are automatically added to.
I have an album shared with my daughter which photos of our dog is automatically added to.
I also like the collages, slideshows, movies and this day x years ago photos which Google Photos automatically creates and notifies me of.
You're willing to pay the price of those algorithms and the Google ecosystem. Others are not.
I'm excited to review this project. Thanks to the creators.
This has come at a perfect moment ... as, this weekend, I'm literally downloading my entire Google photos archive (one year at a time) to my local harddrive and figuring out a way forward.
I'm done with Google after a 'straw breaking the camels back' moment with their payment system.
Why not use Takeout to download all at once?
For me the features that make Google photo, Google photo are:
* it's free and comes by default with an Android phone.
* it just works.
If you can make an effortless way to get online backups of my photos at a reasonable price while regaining privacy, then I'll switch in a heartbeat without a single thought about any of those ML-based moat features Google has crammed in their service.
I want none of those features.
I want automatic backup, easy sharing, and accessibility from all devices.
Personally I'd find the pure storage and basic categories suitable. I dislike almost all the algorithms. Especially "memories" and shit like that.
Simple and reliable backup and reasonably speedy browsing is what I need.
> If I wanted to just store my photos I'd throw them in a S3 bucket or Dropbox or something.
Neither of those give you any privacy unless you do the encryption yourself in which case you have to build something to access them unencrypted. Have you checked out what the service actually does?
Wouldn't a mega encripted folder make sense for the average person?
Let's say you store your photos in Dropbox but inside an encrypted folder. What would you have to do to view the photos? Unless there client you encrypt your files with has a photo viewer, you'd have to download the pictures and decrypt them to look at them. The whole thing becomes very inconvenient very quickly.
1 reply →
On top of this, good algorithms should be run if it is possible to do it in a privacy friendly way.
>So you're going to implement algorithms then?
Jeesh, that's easy.
You encrypt the algorithms too.
I don't want Google at all in my life, so I think this product seems very attractive. But of course it depends on the user, what they value.