News for hackers. Stuff that matters.
Re:Learn TensorFlow and deep learning, without a Ph.D. by amelius on Tuesday January 24, 2017 @08:00AM (#13470943)
Perhaps somebody here can help me with a sideproject that I'm working on. I'm trying to figure out the topology of a neural network that is capable of detecting the location and orientation of a given object. Say, a wrench. I don't want to use heatmaps (e.g. [1]) because they give just the location of the object and not the orientation. So the problem is basically how to choose the output quantities and how to encode them. The x and y coordinates of the head of the wrench could be quantities, but how to encode them? Should I use multiple output neurons per coordinate? And encoding the orientation is a similar problem. Would it even make sense to decompose the output in this way? Thanks in advance!

PS: More generally, is there a guide that explains how to robustly encode real numbers as output of neurons? I've tried to search for it, but couldn't find it.

[1] https://github.com/heuritech/convnets-keras

Re:Learn TensorFlow and deep learning, without a Ph.D. by pizza on Tuesday January 24, 2017 @08:20AM (#13471070)
There are algorithms for stuff just like that in OpenCV. Maybe you could find some inspiration or clarification by reading through the source code for the algorithms for a brief description?
Re:Learn TensorFlow and deep learning, without a Ph.D. by amelius on Tuesday January 24, 2017 @10:15AM (#13471925)
Yes, that's a good idea, thanks!

But I was hoping for a more scientific answer. Like how do researchers approach this problem typically? And is there a strong consensus in this area among researchers?

It seems like such a general problem.

Re:Learn TensorFlow and deep learning, without a Ph.D. by pizza on Tuesday January 24, 2017 @07:30PM (#13477250)
Hm, there are problems in the field of computer vision that might help structure your cost/training algorithm; maybe pose estimation, the perspective-n-point problem, and point-set registration.