At Connect, we shared a look at what we’re calling Instant Codec Avatars—photorealistic representations of yourself created from some scans from your phone + a few hours of processing time. Dive into our Reality Labs Research updates on Tech at Meta: https://t.co/crUTuFeDL4
📢Looking for real-time drivable volumetric avatars? Check out out #SIGGRAPH 2022 paper, DVA!!
We extend Mixture of Volumetric Primitives [Lombardi+ 2021] to articulated human bodies to enable real-time/high-fidelity rendering of clothed human!
PDF: https://t.co/hMYgHzuIiE
1/n Some really mind blowing technological advancements got lost in the excitement over the Meta name change yesterday. Let's start with lifelike avatars - this is an animated avatar not a video recording...
Great work at @CVPR from colleagues from @nc_shape and @EPFL on modulating the correlation between dropout patterns to get a model between dropout and ensemble of independent models.
https://t.co/ERAwpUzOP7
Excited about volumetric avatars and NeRF? 😆
Compositional NeRF for high-quality face rendering has been accepted to CVPR 2021 as an Oral paper.
https://t.co/TgCVNMvEHt
Our upcoming #CVPR2021 paper on efficient uncertainty estimation. tldr: a drop-in replacement for mc-dropout that works (almost) as good as ensembles 😀
very excited to share our recent CVPR 2021 paper "Masksembles for Uncertainty Estimation" where we develop a cheap method to estimate uncertainties for DL models.
co-authors: @psycharo@PierreBaqu1@FuaPv
🕸️project: https://t.co/xYxJ49oMvX
📃paper: https://t.co/2FSw44qG5a
1/7
In practice, standard "deep ensembles" of independently trained models provides a relatively compelling Bayesian model average. This point is often overlooked because we are used to viewing Bayesian methods as sampling from some (approximate) posterior... 1/10
The bike optimized with our deep-learning aerodynamic simulation models was used to beat the world speed record in the women category and the European speed record in the men category!
@nc_shape@EPFL_en@Idiap_ch#DeepLearning
https://t.co/fOXZT7avHL
@ChrisChoy208 We can, but the issue always remains how to efficiently compute the neighbours. I think in this context permutohedral essentially acts as a particular way to hash the inputs. An alternative would be to do sampling or something like locality sensitive hashing?