CVPR 2019

About three months ago, I hit one of my first career goals – have a paper accepted to CVPR (achievement unlocked ! ). Last week I had the privilege to present a poster of our work on normal estimation for unstructured 3D point clouds using CNNs at CVPR 2019 in Long Beach California (16-20.6.19).

In this post, I will summarize my experience from the conference and highlight some papers and workshops that I found interesting (mostly 3D point cloud-related topics but also some nice applications, ideas, and datasets).

If you missed it, the oral presentations are available on this YouTube playlist I composed. All of the computer vision foundation videos are available here.


Get ready for some numbers that will blow you away…
This year CVPR hosted ~9200 attendees from 68 countries. 134 of them are from my home country of Israel (an amazing number considering Israel’s size of the population, as a reference, Germany had 265). In addition, 14104 Authors submitted 5160 papers, of which 1294 were accepted.  If this exponential growth will continue they are expecting 10.8 Billion papers by 2028 (not really).
For all the detail see the IEEE tech news post.


I think the most interesting award is the PAMI Longuet-Higgins Prize “retrospective most impactful paper from CVPR 2009” which was awarded to a paper that achieved the “test of time”. Not very surprisingly, it was given to:

“ImageNet: A large-scale hierarchical image database” by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei.
Very well deserved. When you think about it, it started a new genre of immensely large dataset papers.

Some Israeli pride – best Paper, Honorable Mention: Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Ce Liu, Bill Freeman, and Noah Snavely for their “Learning the Depths of Moving People by Watching Frozen People.” paper. (The paper is not from Israel, but Tali Dekel is… or was… Rumor says she is coming back soon… )

Best Paper: Shumian Xin, Sotiris Nousias, Kyros Kutulakos, Aswin Sankaranarayanan, Srinivasa G. Narasimhan, and Ioannis Gkioulekas for their “A Theory of Fermat Paths for Non-Line-of-Sight Shape Reconstruction.” From Carnegie Mellon University, University of Toronto, and University College London.


In the first two days I attended some great workshops that hosted top tier speakers:

It included one of my favorite spotlight posters of a Eurographics honorable mention paper: Learning a Generative Model for Multi-Step Human-Object Interactions from Videos. He Wang*, Soeren Pirk*, Ersin Yumer, Vladimir Kim, Ozan Sener, Srinath Sridhar, Leonidas J. Guibas.
A talk by Vladlan Koltun provided some interesting empirical evidence to what single-view reconstruction networks learn. Clearly stating that they perform retrieval rather than reconstruction and presented their Oracle NN to support this claim.

Probably the best workshop at the conference. It was so good that I wasn’t able to attend since the room was so full there wasn’t even room to stand. I heard that Prof. Bill Freeman told the computer vision story using rock songs.

Capsule networks are getting increasing attention lately so I was curious what the fuss is all about. After the tutorial, I heard some discussions questioning the effectiveness of Capsule Networks. Geoffrey Hinton said in one of his recent interviews that this is the most interesting topic he is working on today. He usually gets it right… only time will tell. It was nice to see that it is also already extended to point cloud processing.

I especially enjoyed Prof. Thomas Funkhouser’s (Princeton) talk about their TossingBot paper and the idea of “Residual Physics”.

Favorite Posters


  1. Structural Relational Reasoning of Point CloudsYueqi DuanYu ZhengJiwen LuJie ZhouQi Tian[pdf] [bibtex]
  2. 3D Point Capsule Networks 
    Yongheng ZhaoTolga BirdalHaowen DengFederico Tombari
    [pdf] [supp] [bibtex]
  3. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation
    Jeong Joon ParkPeter FlorenceJulian StraubRichard NewcombeSteven Lovegrove
    [pdf] [supp] [bibtex]
  4. PointRCNN: 3D Object Proposal Generation and Detection From Point Cloud
    Shaoshuai ShiXiaogang WangHongsheng Li
    [pdf] [bibtex]
  5. Scan2CAD: Learning CAD Model Alignment in RGB-D Scans
    Armen AvetisyanManuel DahnertAngela DaiManolis SavvaAngel X. ChangMatthias Niessner
    [pdf] [supp] [bibtex]
  6. Learning to Sample
    Oren DovratItai LangShai Avidan
    [pdf] [supp] [bibtex]
  7. Shape Unicode: A Unified Shape Representation
    Sanjeev MuralikrishnanVladimir G. KimMatthew FisherSiddhartha Chaudhuri
    [pdf] [supp] [bibtex]

The last two (6,7) get an “honorable mention” from me for that day. The day ended with a great party by IntelAI.

Day 2

  1. Learning the Depths of Moving People by Watching Frozen People
    Zhengqi LiTali DekelForrester ColeRichard TuckerNoah SnavelyCe LiuWilliam T. Freeman
    [pdf] [supp] [bibtex]
  2. Occupancy Networks: Learning 3D Reconstruction in Function Space
    Lars MeschederMichael OechsleMichael NiemeyerSebastian NowozinAndreas Geiger
    [pdf] [supp] [bibtex]
  3. 3D-SIS: 3D Semantic Instance Segmentation of RGB-D ScansJi HouAngela DaiMatthias Niessner
    [pdf] [supp] [bibtex]
  4. The Perfect Match: 3D Point Cloud Matching With Smoothed Densities
    Zan GojcicCaifa ZhouJan D. WegnerAndreas Wieser
    [pdf] [supp] [bibtex]
  5. GeoNet: Deep Geodesic Networks for Point Cloud Analysis
    Tong HeHaibin HuangLi YiYuqian ZhouChihao WuJue WangStefano Soatto
    [pdf] [bibtex]
  6. PointNetLK: Robust & Efficient Point Cloud Registration Using PointNet
    Yasuhiro AokiHunter GoforthRangaprasad Arun SrivatsanSimon Lucey
    [pdf] [supp] [bibtex]

This day’s honorable mention goes to “The perfect match” paper. Their initial preprocessing of the point cloud reminds me of our 3DmFV representation (though it is very different) and also shows that end-to-end learning is not always the best for point clouds. The day ended with a nice reception with a live band and some state of the … art.

Day 3

I presented my poster today! ! ! Therefore I did not have the chance to see most of the posters in the session. However I did get some sneak peaks before the session started.

  1. Nesti-Net: Normal Estimation for Unstructured 3D Point Clouds Using Convolutional Neural Networks
    Yizhak Ben-ShabatMichael LindenbaumAnath Fischer
    [pdf] [supp] [bibtex]
  2. Generating 3D Adversarial Point CloudsChong XiangCharles R. QiBo Li
    [pdf] [supp] [bibtex]
  3. WarpGAN: Automatic Caricature Generation
    Yichun ShiDebayan DebAnil K. Jain
    [pdf] [supp] [bibtex]
  4. Argoverse: 3D Tracking and Forecasting With Rich Maps
    Ming-Fang ChangJohn LambertPatsorn SangkloyJagjeet SinghSlawomir BakAndrew HartnettDe WangPeter CarrSimon LuceyDeva RamananJames Hays
    [pdf] [supp] [bibtex]


It was an overall amazing experience. The sheer magnitude of the event definitely sets some new standards. I came back to Israel highly motivated and full of ideas. I would like to express my appreciation and gratitude to the conference organizers. Small constructive criticism – please reconsider the oral session format (I realize the time constraints are important but the video and unsynced speakers were less than optimal).

It was also a great chance to meet all of my friends and former colleagues from TUM and ANU, and also make some new ones from all around.
Hope to see you next year!