4 ms·

How Jetpac Built a Photo Quality Algorithm for $5k in 3 Weeks

41 points by datageek 15y ago

waitwhat 15y ago
Spoiler: The algorithms actually work by doing text analysis on the captions and other metadata, and no actual image analysis.
- datageek 15y ago
  the algorithm needed to execute quickly. image analysis takes too long.
  - kabdib 15y ago
    What's the bottleneck? Bandwidth for downloading images? Understanding of an algorithm that would do the job?
    - marshallp 15y ago
      Difficulty in creating an algorithm. There are ways to get it done algorithmically, however, the challenge is in getting enough data. 30,000 images is too low. you would need a few million, then just simple machine learning algorithms would work. The latest machine learning techniques such as unsupervised deep learning might work, with millions of unlabeled images and the 30,000 labelled.
    - petewarden 15y ago
      Bandwidth and general cumbersomeness of dealing with larger amounts of data with starving-startup resources. I actually spent about a decade of my career focused on image processing, and while I love it's power, I knew how much of an engineering challenge it can be at massive scale. I need to do a blog post about this, since I know my choice is a bit surprising and needs explanation.
  - lwat 15y ago
    Technically I wouldn't call it a 'Photo quality algorithm' in that case
- Drbble 15y ago
  Spoiler 2: They achieved the low cost by exploiting Kaggle's army of machine learning practitioner suckers/volunteers/competitors.
- socialist_coder 15y ago
  Pretty disappointing. The image analysis was the only reason I clicked.
brown9-2 15y ago
Is 30,000 photos a large enough dataset for a ML exercise like this? It might just be journalistic simplifying but I wonder if they overfit their data with such specific "good" and "bad" words: Among the best words: Peru, Cambodia, Michigan, tombs, trails and boats. What photo captions are the most likely to signify a bad photo for a travel magazine? San Jose, mommy, graduation and CEO, Warden says.
humbledrone 15y ago
The 60 second challenge thing strikes me as extremely short-sighted. It is impressive to be able to solve a series of tiny bugs with just 60 seconds for each one, but I don't think it's the right metric to find the elusive "10x" programmer. It's my opinion that "10x" programmers are an order of magnitude faster than their peers not because they are lightning-fast at spitting out code (although they may be). They're faster because they can analyze the problem, drill down to its essence, and come up with an effective plan for solving it. Contrast this with a poor programmer, who might code circles around the problem, wasting time writing code that could just be taken from off the shelf. It doesn't matter if the poor programmer somehow is brilliant at punching out correct code if the code they're punching out didn't need to be written in the first place. It's like selecting creative writers by holding a contest to see which ones can spot a grammatical mistake the most quickly. Sure, that's a helpful skill for a writer to have, but it doesn't mean that they'll write anything worth reading.
- kkowalczyk 15y ago
  Or we can assume that the company that does the hiring via 60 second challenge is not ran by complete idiots and they do keep track whether those hires end up being effective employees for them. Let's give them the benefit of a doubt. Your argument seems to be that there are only two possible programmers: effective designers that are slow to code or fast coders that are bad designers (designer in terms of code architecture, not graphic design, of course). I claim that there are no good architects that are slow coders i.e. all good architects are also good coders. When you look into research of skill acquisition, or even common sense, in order to become good at something you first have to master the basics before you master higher-level skills. In the field of programming, being able to implement a piece of highly constrained code are the basics of our craft. Once you're really good at that, people move up to tasks that have less constrains and are therefore high-level: choosing an implementation language, designing how pieces of the whole app fit together etc. But you won't ever get to do that if you were not able to quickly and confidently implement a function that someone else specified, just like there are no chefs who can't cut vegetables very quickly or basketball players that are brilliant tacticians but can't run very fast etc. Mastery of basics is not a guarantee of mastery of higher-level skills but it is a prerequisite.
Iroiso 15y ago
The essence of the post is about using crowd sourcing to alleviate start-up challenges, I think they did a good job of exposing these opportunities to founders, (BTW, its also great PR for kaggle), Good Article all in all.
- coderdude 15y ago
  I'm not sure why you were down-voted at least a couple times for this comment. You've done a much better job of summarizing what this article was about than the "spoiler:" above (which is more based on the title than the point of the article).

41 points by datageek 15y ago

Spoiler: The algorithms actually work by doing text analysis on the captions and other metadata, and no actual image analysis.

datageek 15y ago
the algorithm needed to execute quickly. image analysis takes too long.
- kabdib 15y ago
  What's the bottleneck? Bandwidth for downloading images? Understanding of an algorithm that would do the job?
  - marshallp 15y ago
    Difficulty in creating an algorithm. There are ways to get it done algorithmically, however, the challenge is in getting enough data. 30,000 images is too low. you would need a few million, then just simple machine learning algorithms would work. The latest machine learning techniques such as unsupervised deep learning might work, with millions of unlabeled images and the 30,000 labelled.
  - petewarden 15y ago
    Bandwidth and general cumbersomeness of dealing with larger amounts of data with starving-startup resources. I actually spent about a decade of my career focused on image processing, and while I love it's power, I knew how much of an engineering challenge it can be at massive scale. I need to do a blog post about this, since I know my choice is a bit surprising and needs explanation.
- lwat 15y ago
  Technically I wouldn't call it a 'Photo quality algorithm' in that case
Drbble 15y ago
Spoiler 2: They achieved the low cost by exploiting Kaggle's army of machine learning practitioner suckers/volunteers/competitors.
socialist_coder 15y ago
Pretty disappointing. The image analysis was the only reason I clicked.

brown9-2 15y ago

Is 30,000 photos a large enough dataset for a ML exercise like this? It might just be journalistic simplifying but I wonder if they overfit their data with such specific "good" and "bad" words: Among the best words: Peru, Cambodia, Michigan, tombs, trails and boats. What photo captions are the most likely to signify a bad photo for a travel magazine? San Jose, mommy, graduation and CEO, Warden says.

humbledrone 15y ago

The 60 second challenge thing strikes me as extremely short-sighted. It is impressive to be able to solve a series of tiny bugs with just 60 seconds for each one, but I don't think it's the right metric to find the elusive "10x" programmer. It's my opinion that "10x" programmers are an order of magnitude faster than their peers not because they are lightning-fast at spitting out code (although they may be). They're faster because they can analyze the problem, drill down to its essence, and come up with an effective plan for solving it. Contrast this with a poor programmer, who might code circles around the problem, wasting time writing code that could just be taken from off the shelf. It doesn't matter if the poor programmer somehow is brilliant at punching out correct code if the code they're punching out didn't need to be written in the first place. It's like selecting creative writers by holding a contest to see which ones can spot a grammatical mistake the most quickly. Sure, that's a helpful skill for a writer to have, but it doesn't mean that they'll write anything worth reading.

kkowalczyk 15y ago
Or we can assume that the company that does the hiring via 60 second challenge is not ran by complete idiots and they do keep track whether those hires end up being effective employees for them. Let's give them the benefit of a doubt. Your argument seems to be that there are only two possible programmers: effective designers that are slow to code or fast coders that are bad designers (designer in terms of code architecture, not graphic design, of course). I claim that there are no good architects that are slow coders i.e. all good architects are also good coders. When you look into research of skill acquisition, or even common sense, in order to become good at something you first have to master the basics before you master higher-level skills. In the field of programming, being able to implement a piece of highly constrained code are the basics of our craft. Once you're really good at that, people move up to tasks that have less constrains and are therefore high-level: choosing an implementation language, designing how pieces of the whole app fit together etc. But you won't ever get to do that if you were not able to quickly and confidently implement a function that someone else specified, just like there are no chefs who can't cut vegetables very quickly or basketball players that are brilliant tacticians but can't run very fast etc. Mastery of basics is not a guarantee of mastery of higher-level skills but it is a prerequisite.

Iroiso 15y ago

The essence of the post is about using crowd sourcing to alleviate start-up challenges, I think they did a good job of exposing these opportunities to founders, (BTW, its also great PR for kaggle), Good Article all in all.

coderdude 15y ago
I'm not sure why you were down-voted at least a couple times for this comment. You've done a much better job of summarizing what this article was about than the "spoiler:" above (which is more based on the title than the point of the article).