Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Cloud Video Intelligence API

cloud.google.com

Post reply on HN
Print view View original

61–70 of 91 posts

Re: Cloud Video Intelligence API

#61

Post by kevin488 »

chrisballinger wrote: post #54
tyre wrote: post #50

I think their model should take a second pass on the words and probabilities, independent of the video. Look at their example: Animal: 97.76% Tiger: 90.11% Terrestrial animal: 68.17% So we are 90% sure it is a tiger but only 68% sure it is a land animal? I don't think that makes sense. It could be that this is a weakness of seeding AI data with human inputs. I can believe that 90% of people who saw the video would ag…

Perhaps it's confused by the many images and videos of tigers swimming in water?

Thanks a lot Life of Pi

Re: Cloud Video Intelligence API

#63

Post by crack-the-code »

tyre wrote: post #50

I think their model should take a second pass on the words and probabilities, independent of the video. Look at their example: Animal: 97.76% Tiger: 90.11% Terrestrial animal: 68.17% So we are 90% sure it is a tiger but only 68% sure it is a land animal? I don't think that makes sense. It could be that this is a weakness of seeding AI data with human inputs. I can believe that 90% of people who saw the video would ag…

It's probably more likely that they want each output to be independent of the other. Certain features may be predominantly associated with a tiger, but not necessarily indicative of a terrestrial animal. If the 9.89% chance that they could have been wrong would have been the case, then that should not influence whether or not it was a terrestrial animal. In my opinion, the consumer of the output values should be able to rely on these fields independently, and make these associations themselves. Although I totally agree a second pass could be useful as a separate data set.

Re: Cloud Video Intelligence API

#64

Post by aub3bhat »

timc3 wrote: post #60
aub3bhat wrote: post #6

I think there is a need for a comprehensive system for image and video data analytics. Much like how we today have relational databases (postgres, MYSQL) and full text search engines (lucene/Solr). The approach Google or Amazon have been taking which involves providing a "tagging" API is frankly unimaginative. I am working on Deep Video Analytics an Open Source Visual Search and Analytics platform for images and vide…

I would be interested to know more about this, particularly the database and what you plan to do with it in the future (I am thinking the license on the GitHub project is obviously restrictive for a purpose at the moment).

Sorry about the license, I am trying to reach a beta version within a month along with a system-description paper that outlines the long term vision behind building such a system. At that point I plan on relaxing the license. There are certain constraints such as making sure that all underlying models are correctly licensed. Also FAISS which I use is licensed by Facebook under an explicit non-commercial license.

Re: Cloud Video Intelligence API

#67

Post by kneel »

sna1l wrote: post #2

I wonder if Snapchat is/will become a large user of this service? Depending on the average response time of this API, Snapchat could get much better ad targeting analyzing their Stories content. I imagine that they have something similar in house that they run since it is pretty vital to their core business, but you never know.

They're more than likely on this already.

Any photo you save in the app is already categorized by content.

Re: Cloud Video Intelligence API

#68

Post by njohnson41 »

crack-the-code wrote: post #63
tyre wrote: post #50

I think their model should take a second pass on the words and probabilities, independent of the video. Look at their example: Animal: 97.76% Tiger: 90.11% Terrestrial animal: 68.17% So we are 90% sure it is a tiger but only 68% sure it is a land animal? I don't think that makes sense. It could be that this is a weakness of seeding AI data with human inputs. I can believe that 90% of people who saw the video would ag…

It's probably more likely that they want each output to be independent of the other. Certain features may be predominantly associated with a tiger, but not necessarily indicative of a terrestrial animal. If the 9.89% chance that they could have been wrong would have been the case, then that should not influence whether or not it was a terrestrial animal. In my opinion, the consumer of the output values should be able…

Still, in any consistent way of assigning probabilities to events, if A implies B, then P(A) Neural network outputs are not probabilities. I think that's the main lesson here.

Re: Cloud Video Intelligence API

#69

Post by cobookman »

tambourine_man wrote: post #49

It amazes me how smart these guys at google are, and yet, they can't design a mobile site if their lives depended on it: http://imgur.com/bXGuNfL

If you could share with me what mobile phone / web browser you used that produced the styling issue, I'll be sure to pass it on to the relavent people within google so that it gets resolved.

Also if you send me an email at bookman@google.com I'll be sure to update you as to when the styling errors are resolved.

(Disclaimer, I work for google cloud)

Re: Cloud Video Intelligence API

#70

Post by imh »

muzakthings wrote: post #44
imh wrote: post #36

The demo picture they chose is interesting. It's obviously a tiger, and is identified as such with only 90% probability. I appreciate the difficulty of the problem and how big of a success it is to achieve even that level of confidence, but that low level of confidence really shows how far we are from being able to simply trust computer vision. Still useful from an information retrieval perspective, I expect.

You realize that softmax scores aren't probabilities, right? It's just a relative measure of confidence, scaled such that they all sum to 1.0.

That's kinda true, but (regularization aside) for standard loss functions it's minimized at the point it's well calibrated, right? Given the scores in the image (97% animal, 90% tiger, etc) they seem to be binary classifiers e.g. "is this a tiger?" So of all scores in the neighborhood of 90%, 90% should be "yes it is," making it a measure of confidence compatible with probability.

Please someone correct me if I'm wrong, but I'm pretty sure that's how it works, just like how logistic regression gives you a probability.

Post reply on HN