Algorithm improvement for Coca-Cola can shape recognition | Hacker News Reader

Algorithm improvement for Coca-Cola can shape recognition

stackoverflow.com

76 pointssun12312 comments

A gorgeous question, but also somewhat controversial.

Corresponding meta discussion, mentioned in the question: http://meta.stackoverflow.com/questions/129423/on-the-bounda.... I merely wonder why it's tagged 'java'.

This computer vision stuff is so fascinating. Does anyone know what the career prospects are like if you specialize in this? Is there much consulting demand? What kind of fees do CV consultants charge?

I have worked in the computer vision field, although never as a consultant, so I can't give you those details. It is used extensively in system automation and manufacturing. The most popular products I've used in that realm were from Cognex. They have some pretty amazing stuff.

I feel like that there is good range of job opportunities in related fields: industrial machine vision (quality control, inspection, measurement), mobile (image enhancement, stabilization), medical, military, mapping, "consumer toys". Even in Finland there is quite many machine vision companies which specialize in designing, developing algorithms, building, coding and installing machine vision systems.

I personally really enjoy working with images and the projects have had good variety even if they have all been about industrial machine vision (measurement). At least in Finland demand for industrial machine vision seems to be good, but "machine vision" has still certain stigma because of previously failed projects. Many projects have failed because of lack of good hardware (cameras, illumination, computing power) and knowledge. At least the hardware side is now in quite good condition.

i think you dont need to be a consultant in this, i mean its pure technique, if your technical and understand math, then you might be able to make something. you need the spirit of hackaday to be working in it, and dont think of goldmines, there is lots of competition. on the positives lots is done at universities (and they lack a lot of practical insight). So there is room for improvement, you might try to build your own coca colla finder just to train yourself.

The original post really highlights the "problem with computer vision" - finding the coke can is fairly easy (SIFT), although the author goes off on some other random approach that I doubt would work very well. The field certainly has its share of Do-It-Yourselfers, plus it's also quite common to outsource labour overseas.

As for fees, it's entirely reflective of a) how much money the problem is costing the client and b) your ability to prove to them your value to the client. Clients do not always respect the complexity of the problems they're facing, so it's easier to work with those who have tried for themselves.

For what it's worth, we're hiring right now in this space (computer vision software development), located in Ottawa, Canada: http://ca.indeed.com/job/software-engineer-71a851096a104223

I don't see any problem with the original solution. It is classic (but little bit old fashioned) way of detecting objects, far from random. I also think that you oversimplify the SIFT solution. You use sift, then what? Match the feature vectors, get bunch of possible correspondence points, prune outliers using 3D model of a cylinder (or assume that the face of the can is flat plane and just use homography) and robust fitting, validate fitting, find other cans, if none found change parameters? No preprocessing? Ignore color? Etc. Not so straight forward.

SIFT is faster than generalized HoughTF in this case, use of points instead of whole objects/blobs gives you nice ways to validate using 3D model, etc., but I would still not look down on some one who finds and implements this solutions as a student.

You're right, it's ok for a project. I guess what I meant was it's quite easy compared to the problems my company deals with every day. 30 images in 24 hours (I think it said) is kind of a problem.

Given the constraints of the asker (all available in OpenCV), the solution I would suggest first is SIFT + homography, both easy to use and in OpenCV (sample code is around). Yes, there's a lot of other possibilities, but this would be an improvement over the original.

Hey that sounds like a cool job but why does the candidate need to know java and c++?

Well we do use both (and various others), but really what languages someone knows isn't such a big factor.

Okay, someone needs to create an open repo of all these recognition algos before Apple or Google patents everything.

NOTE: SURF and SIFT are both already patented, btw, along with some other things in OpenCV, so watch out with them.

Reply on news.ycombinator.com