20211114-20211127.md

意见性分享

论文公式排版

Kosta Derpanis

Have a few spare moments before the #CVPR2022 submission deadline? Check out this great guide to making appealing LaTeX tables: https://people.inf.ethz.ch/markusp/teaching/guides/guide-tables.pdf

• Erfan: I found https://tablesgenerator.com useful.

reference: https://twitter.com/CSProfKGD/status/1460345329895841797?s=20

论文写作小建议

Kosta Derpanis

Great collection of technical writing pet peeves by @tamaramunzner! https://www.cs.ubc.ca/~tmm/writing.html

• Tamara Munzner: The crusade where I’m so often shouting into the wind is my hatred of the solo / non-referential “This”: https://www.cs.ubc.ca/~tmm/writing.html#this

reference: https://twitter.com/CSProfKGD/status/1460485971317768193?s=20

临近投稿时间的建议

Kosta Derpanis

#CVPR2022 tip: Make sure you periodically download a local copy of your LaTeX files. In the past, online services have struggled to stay up near the submission deadline.

• Michael Black: Better yet, pay for the upgraded version of Overleaf and you can automatically sync it with your Dropbox. Then if Overleaf goes down, you have a version on your disk. Magic. At least one person on every team should do this. Tip: if someone just shares the link with you, it doesn't sync with Dropbox. You need to "join" the project.

reference: https://twitter.com/CSProfKGD/status/1460608992808517632?s=20

课程和报告分享

---

成果推荐及讨论

• ML with differential geometry/algebraic topology

  Michael Bronstein

  After a hiatus, a new series of blogs posts. Do differential geometry and algebraic topology sound too exotic for ML? In recent works, we show that tools from these fields bring a new perspective on graph neural networks.

  First post in the series: https://towardsdatascience.com/graph-neural-networks-through-the-lens-of-differential-geometry-and-algebraic-topology-3a7c3c22d5f

  Part II will discuss topological message passing. Part III will deal with geometric flows and non-euclidean diffusion PDEs on graphs. Part IV will show how the over-squashing phenomena relate to graph curvature, and offer a graph rewiring approach inspired by the Ricci flow.

  Beltrami flow (BLEND) #NeurIPS2021 paper: https://arxiv.org/abs/2110.09443

  GRAND #ICML2021 paper: https://arxiv.org/abs/2106.10934

  Topological message passing #ICML2021 and #NeurIPS2021 papers: https://arxiv.org/abs/2103.03212, https://arxiv.org/abs/2106.12575

  reference: https://twitter.com/mmbronstein/status/1461366066282569734?s=20

• self-supervised multilingual model

  Meta AI

  We're releasing XLS-R, a self-supervised multilingual model trained on 128 languages for a variety of speech tasks. XLS-R sets a new state of the art on several speech recognition benchmarks, foreign-to-English speech translation & language identification. https://ai.facebook.com/blog/xls-r-self-supervised-speech-processing-for-128-languages

  reference: https://twitter.com/MetaAI/status/1461375337174945793?s=20

• scan with iPhone 13 Pro

  EveryPoint

  We modeled the 11-foot-8 Bridge (aka The Can Opener) in Durham, NC using JUST an iPhone 13 Pro. With appropriate scale control in our scan scenes, we can scan overpasses with dimensional accuracy using consumer grade cameras.

  reference: https://twitter.com/EveryPointIO/status/1461488874409005056?s=20

• LOLNeRF

  AK

  LOLNeRF: Learn from One Look

  abs: https://arxiv.org/abs/2111.09996

  a method for learning a generative 3D model based on neural radiance fields, trained solely from data with only single views of each object

  reference: https://twitter.com/ak92501/status/1462597053180223493?s=20

• TermiNeRF

  Ronnie Clark

  Check out TermiNeRF (to appear at #3DV2021)! We significantly improve rendering speed (~14x faster than NeRF) by learning ray -> depth predictions. We show this allows fast finetuning to new lighting conditions. arxiv.org/abs/2111.03643

  reference: https://twitter.com/ronnieclark__/status/1463546853119729668?s=20

• quickly training NeRF-like models

  Jon Barron

  Some great progress on quickly training NeRF-like models made mostly of voxel grids. 49x faster training than mip-NeRF, wow! https://arxiv.org/abs/2111.11215

  

  reference: https://twitter.com/jon_barron/status/1463649912134922248?s=20

• NÜWA

  AK

  NÜWA: Visual Synthesis Pre-training for Neural visUal World creAtion presents a unified multimodal pretrained model that can generate new or manipulate existing visual data (i.e., images and videos) for various visual synthesis tasks.

  

  reference: https://twitter.com/ak92501/status/1463691106416214019?s=20

• 3D Models, Materials, and Lighting From Images

  AK

  Extracting Triangular 3D Models, Materials, and Lighting From Images present an efficient method for joint optimization of topology, materials and lighting from multi-view image observations.

  abs: https://arxiv.org/abs/2111.12503

  

  reference: https://twitter.com/ak92501/status/1463703734815539209?s=20