top | item 17841819

Beyond the pixel plane: sensing and learning in 3D

129 points| gtmtg | 7 years ago |thegradient.pub | reply

15 comments

order
[+] JeremyHerrman|7 years ago|reply
This is an exciting time for those of us working on computational geometry to better understand 3D shapes across many industries.

In addition to the architectures mentioned in this great overview, I'm excited to see progress on spectral and geodesic CNNs for graphs and manifolds. Check out this other fantastic source for info on 3D ML: http://geometricdeeplearning.com

[+] jeffchuber|7 years ago|reply
This is a great overview. Also checkout CS 468 from Stanford, http://graphics.stanford.edu/courses/cs468-17-spring/ "Machine Learning for 3D Data"

Also, if you want to work on this stuff full time- https://news.ycombinator.com/item?id=17649726

[+] bitL|7 years ago|reply
Super cool course! Thanks for the link!

Do you know what are the most precise programmable RGB-D cameras a non-professional can buy? I was trying to extract 3D information just from a single camera via 3D convolutions and RNNs (for a self-driving car project) and would like to play with real 3D a bit as well.

[+] glalonde|7 years ago|reply
What about pose estimation? e.g. Given a well defined coordinate system, like the origin is the nose on a face, determine the pose of the face. Is this still best done with classic optimization formulations like ransac/ICP and a supplied model, or have these been bested by learned models somehow?
[+] gtmtg|7 years ago|reply
Don't think it's exactly what you're talking about (I'm sure there are other works much closer to what you have in mind, just can't recall off the top of my head) — but you might find PoseNet (https://www.cv-foundation.org/openaccess/content_iccv_2015/p...) interesting. Not explicitly 3D, but estimates where in a large-scale scene a picture was taken using an end-to-end convolutional network.

With that said, I think there's still a ton of merit in classical geometric approaches like ICP — there's a real, geometric basis to why they work. Convolutional networks can demonstrate some pretty amazing results, but they're still mostly "black boxes" to us, and a consequence of this is that it's hard to understand why they work and predict when they'll fail. This blog post (by the PoseNet author, actually) articulates the viewpoint well: https://alexgkendall.com/computer_vision/have_we_forgotten_a.... One recent research direction that I personally find really fascinating is designing deep learning architectures around real geometric properties, e.g. as in Skydio's deep stereo work: https://arxiv.org/pdf/1703.04309.pdf

[+] ajmarcic|7 years ago|reply
Has any progress been made towards single view 2D -> 3D inference?