SLIDE 1
CVPR is a contemporary art exhibition
- - Garbage is a source for impact --
3D Scene Understanding for Vision, Graphics, and Robotics Workshop @ CVPR2020
Yasutaka Furukawa
CVPR is a contemporary art exhibition -- Garbage is a source for - - PowerPoint PPT Presentation
CVPR is a contemporary art exhibition -- Garbage is a source for - - PowerPoint PPT Presentation
CVPR is a contemporary art exhibition -- Garbage is a source for impact -- 3D Scene Understanding for Vision, Graphics, and Robotics Workshop @ CVPR2020 Yasutaka Furukawa We are scared of reviewers HARSH REVIEWER CVPR is a contemporary art
SLIDE 2
SLIDE 3
CVPR is a contemporary art exhibition
- - Garbage is a source for impact --
SLIDE 4
SLIDE 5
5
CAD model Structured geometry Semantically segmented geometry
- Model editing
- texture/normal/bump/effects mapping
- Physics simulations
- Interactions (VR/AR)
- Architectural analysis
- …
[ Transformer ]
SLIDE 6
6
SLIDE 7
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Structured Geometry Reconstruction
SLIDE 8
SLIDE 9
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age Revolution
SLIDE 10
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
SLIDE 11
SLIDE 12
SLIDE 13
SLIDE 14
CVPR is a
SLIDE 15
SLIDE 16
CVPR is a contemporary art exhibition Let’s stop pretending that all our 2000 CVPR papers convey good ideas Garbage is a source for impact
SLIDE 17
We are scared of reviewers…
HARSH REVIEWER
SLIDE 18
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
SLIDE 19
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
SLIDE 20
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
Manhattan World Stereo (387 citations) [ Furukawa, Curless, Seitz, and Szeliski ]
SLIDE 21
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
SLIDE 22
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
Chen Liu
Research idea-flow
SLIDE 23
Research idea-flow
CVPR 2009 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
Chen Liu
ECCV 2012
Top-down is better than bottom-up Detect higher-level primitive
SLIDE 24
Detect higher-level primitives
3D points Floorplan
SLIDE 25
Bottom-up: Corner Detection
3D points Floorplan
SLIDE 26
Bottom-up: Corner Detection
One missing corner
SLIDE 27
Bottom-up: Corner Detection
One missing corner
- 1. Cannot miss a single primitive.
- 2. Enumerate much more than necessary.
- 3. # of primitives to enumerate grow exponentially as the primitive DoF grows.
SLIDE 28
Top-down: Region Detection
SLIDE 29
Top-down: Region Detection
SLIDE 30
Research idea-flow
CVPR 2009 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
Chen Liu
ECCV 2012
Top-down is better than bottom-up Detect higher-level primitive
SLIDE 31
Research idea-flow
CVPR 2009 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
Chen Liu
ECCV 2012 ECCV 2018
In a top-down approach, shape-grammar is a restrictive shape prior
SLIDE 32
Shape grammar as a shape prior
SLIDE 33
40
SLIDE 34
Grammar cannot model complex buildings
SLIDE 35
Inverse Procedural Modeling Manual design
Shape grammar
Procedural Reconstruction
SLIDE 36
Inverse Procedural Modeling Manual design
Shape grammar
Procedural Reconstruction
Big-data Machine-learning Technique
Maybe a good idea
SLIDE 37
Research idea-flow
CVPR 2009 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
Chen Liu
ECCV 2012 ECCV 2018
In a top-down approach, shape-grammar is a restrictive shape prior
SLIDE 38
Research idea-flow
CVPR 2009 ICCV 2015 CVPR 2014 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
ECCV 2012
DNN enables the use of
- ptimization instead of replacing
ICCV 2017
Chen Liu
SLIDE 39
Google Maps - satellite (2D) CAD model / Architect drawing
Outdoor Architecture Vectorization
SLIDE 40
Corners Edges Regions
SLIDE 41
Corners Edges Regions
Optimization (Integer programming)
SLIDE 42
- DNN shrinks candidate pool
- DNN make energy landscape easy
DNN enables optimization instead of replacing
Scalability Inference power MRF A few million vars Pairwise Mostly submodular IP A few hundred vars Higher-order Non-submodular
SLIDE 43
Research idea-flow
CVPR 2009 ICCV 2015 CVPR 2014 ECCV 2018 ECCV 2018/ICCV 2019
Ice-age (no impact) Revolution (impact)
ECCV 2012 ICCV 2017
Chen Liu
CVPR 2020
Instance segment works for geometric segments
CVPR 2018/CVPR 2019
SLIDE 44
Research idea-flow
CVPR 2009 ICCV 2015 CVPR 2014 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
ECCV 2012 ICCV 2017
Chen Liu
- 1. Neural topological inference
(avoid complex optimization)
- 2. Implicit grammar learning
CVPR 2020
SLIDE 45
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
Chen Liu
Research idea-flow
SLIDE 46
Research idea-flow
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
SLIDE 47
Research idea-flow
CVPR 2009 ECCV 2012 ICCV 2015 CVPR 2014 ICCV 2017 ECCV 2018 ECCV 2018/ICCV 2019 CVPR 2020 CVPR 2018/CVPR 2019
Ice-age (no impact) Revolution (impact)
SLIDE 48