ๆญฃๅจๅ ่ฝฝ่ง้ข...
่ง้ขๅ ่ฝฝๅคฑ่ดฅ
๐ข๐ข๐ข "๐๐๐๐ข๐๐ง๐ญ ๐ ๐จ๐๐ฆ: Real-Time Differentiable Ray Tracing", a mesh-based 3D represention. Co-lead by my PhD students Shrisudhan Govindarajan and Daniel Rebain, and w/ Kwang Moo Yi
115,036 ๆฌก่ง็ โข 1 ๅนดๅ โขvia X (Twitter)
12 ๆก่ฏ่ฎบ
The core idea is to use a 3D polygonal mesh to perform differentiable optimization... but how to differentiate wrt. its combinatorial structure? Turns out the Voronoi diagram is differentiable w.r.t. the positions of its sites ๐คฉ(while Delaunay triangulations are not ๐งโ๏ธ)
And how do you render/train this? โย sorting tricks? Nope. โย rely on Ray tracing libraries like Optix? No way!! Just "walk" along the Voronoi cells! (you can think this as a generalization of a Digital Differential Analyzer, or DDA, to convex polygonal domains)
And how does it perform? Very competitively, but the biggest gain is how much simpler everything becomes (after some gruesome CUDA coding ๐)
Let me wrap up by saying how proud I am of Daniel and Shri โค๏ธโ๐ฅโค๏ธโ๐ฅโค๏ธโ๐ฅ. This was a much bigger project than the vast majority I have seen in recent times... and it took MONTHS of extreme hard work, dedication, and resilience (yes, this is really the tip of the iceberg). Enjoy the paper!
Source Code is out in a few days by the wayโฆ. Mostly readme/install and alpha testing it with an external party. Thanks @janusch_patas
๐ How is #AI changing global #logistics? Listen to this episode with @ScottWLuton and Jake Barr with Hans Moerman from @DP_World discussing the role of digital transformation in shaping the future of supply chains. ๐ง Listen now:
@kwangmoo_yi Looks super nice! But it seems that the statement โDelaunay Triangulations are not differentiable w.r.t. to positions of its sitesโ is not correct, as we proposed the very idea in our paper: ๐ฎ Can you please check it?
@kwangmoo_yi Wow, this is a really cool idea!
@kwangmoo_yi Iโm curious how you guys managed to get such high frame rates with a raytraced algorithm! Isnโt the thread divergence high if the threads in a block arenโt looking at the same primitive? Itโs also a lot of uncoaleased reads. How long until we get more details?
@kwangmoo_yi Details? About a week when code comes out? Is there an experiment you are suggesting?
@kwangmoo_yi Great work, how much improvement can we expect from ray tracing on a volume followed by marching cube to generate a mesh, compared to ray tracing on a Voronoi diagram followed by Delaunay triangulation to generate a mesh?
@kwangmoo_yi Marching tets is usually superior in terms of quality. Not bounded by Nyquist in terms of max frequency
