Skip to content

Latest commit

 

History

History
62 lines (58 loc) · 6.8 KB

20220502-20220515.md

File metadata and controls

62 lines (58 loc) · 6.8 KB

成果推荐及讨论

  • Stephen James
  • Srinath Sridhar
    • CVPR work(ConDor):How can we train our neural networks to make 3D chairs, tables, and other objects upright? In our upcoming work we investigate the problem of "canonicalizing" the 3D pose of common object categories without any supervision.
    • ConDor uses Tensor Field Networks (TFNs - neural networks that are equivariant to point permutation, 3D rotation, and translation) to estimate a canonical frame of reference that is learned using self-supervision losses on common shape categories. ConDor can also handle partial 3D shapes as shown below. Surprisingly, it can also consistently co-segment shapes without any supervision.
    • Project Webpage: ivl.cs.brown.edu/ConDor/
  • Xiaohua Zhai
    • We release the Big Vision codebase, a JAX library originally used to develop ViT, Mixer, ViT-G, LiT, and more! Together, a better plain ViT-S/16 baseline (76.5% ImageNet, 90 epochs) is provided, as a simple and strong starting point. We are thrilled to announce the Big Vision codebase that supports training large-scale vision models on Google Cloud TPUs. It scales seamlessly from a single core to up to 2048 cores!
    • github
  • Daqi Lin
    • Want real-time global illumination beyond diffuse? Introducing ReSTIR Path Tracing (ReSTIR PT) that allows you to reuse paths through glass and other complex interactions, based on a new theory we develop - Generalized Resampled Importance Sampling.
    • github
  • Zhiqin Chen
    • Announcing Neural Dual Contouring (NDC), a new data-driven approach to reconstructing meshes from all kinds of inputs: grids of signed or unsigned distances, binary voxels, or point clouds (without normals). Compared to our prior work Neural Marching Cubes, it is simpler, faster, more robust, and able to take unsigned inputs.
    • github
    • Andrea Tagliasacchi:The network is VERY simple, given a multitude of input formats use a neural network to regress:
      1. polygon existence on facets (... just a grid)
      2. vertex coordinates within cells (... just a grid) Then you stitch everything up with the classical dual contouring logic... Voilà!
  • Zirui Wang
    • CoCa: a new image-text foundation model subsuming single-encoder, dual-encoder and encoder-decoder. SOTA results on 19 unimodal/multimodal/alignment tasks including 86.3% zero-shot top-1 ImageNet, 90.6% with a frozen encoder, 91.0% when finetuned.
    • link
  • Iliyan Georgiev
  • AK
  • Rana Hanocka
  • Xiaowei Zhou
  • Wenzel Jakob

课程和报告分享/课程和会议预告

意见分享及讨论

工具分享

招聘

  • We are looking for a PostDoc at the Computer Vision and Geometry Group (CVG) at ETH Zürich. The candidate should have strong expertise in 3D vision and/or mobile robotics and have papers published at top-tier ML, robotics, or compu -link
  • Regular reminder that Qualcomm AI Research is hiring DL researchers and software engineers! -link