Computers are actually better at object recognition than humans now.
11 plus ones
Shared publicly•View activity
- I wonder if humans would win for a stereo image test set.Feb 21, 2015
- For object recognition or depth perception or what?Feb 21, 2015
- Even if computers are too good at finding cat pictures now, we humans can probably still beat them on finding cute cat pictures. We are really good at dealing with ill-defined problems; computers? not so much. :)Feb 21, 2015
- Good point, jilin. That's a good point that deep learning might have trouble with humor.Feb 21, 2015
- Object recognition. Stereo imagery has more information. Of course, NNs could be trained with stereo image data, which has been done before, but there isn't as much training data available.Feb 22, 2015
- If computers are already beating humans using a single image, then in the stereo case, just have the algorithm run on one image instead of two, and it will probably still beat out humans. Object recognition generally don't rely on stereo vision very much.Feb 22, 2015
- Funny, G+ is not sending me email notifications on this thread...
Anyway, to avoid a lengthy reply, I'll just try to list a few thoughts:
• I'm not a NN expert but have gone through the USPO handwritten digit NN exercise in Prof. Ng's ML course.
• My point was that I feel humans might score better if presented with stereo imagery.
• Human vision samples multiple points on an object (between saccades) and the brain's depth perception allows it to build a relational graph between these points (example predicates: RelativeDepth, Texture, etc.) and this additional information helps the brain classify an object (in a stereo pair).
• I feel that using RGB training data for NNs alone (if this is the case) may be shortsighted.
• I think the size/performance of the deep/massive image recognition NNs could be reduced/improved, respectively, if input data besides RGB training data were used (stereo or disparity map, object class 3D wire meshes, FFT, others).
Thanks for sharing this post, and have a nice Sunday :-)Feb 22, 2015
- made the interesting claim on a different thread that the errors that the humans and computers make appear to be quite different. Apparently, the computer tends to miss broad categories (identifying a dog as a sheep, for example), but the humans tend to miss narrow categories (identifying a dog as a dog, but getting the dog breed wrong).
If that's correct, it's interesting in a few ways. First, the computer errors are probably worse than the score would indicate, at least in real world applications. Second, it probably means the computer is prioritizing some features humans consider less important, and ferreting out those differences might be interesting, as would tuning the computer system to weight broad category errors as more severe. Third, it suggests there still could be value in combining humans and computers (such as kicking out the detection to MTurk when the computer is uncertain).Feb 22, 2015
Add a comment...