Automatic 3D modeling from range images Daniel Huber Carnegie Mellon University Robotics Institute.

Slides:



Advertisements
Similar presentations
Image Registration  Mapping of Evolution. Registration Goals Assume the correspondences are known Find such f() and g() such that the images are best.
Advertisements

 Over-all: Very good idea to use more than one source. Good motivation (use of graphics). Good use of simplified, loosely defined -- but intuitive --
The fundamental matrix F
Edge Preserving Image Restoration using L1 norm
Surface Simplification Using Quadric Error Metrics Speaker: Fengwei Zhang September
Presented by Konstantinos Georgiadis. Abstract This method extends the Hierarchical Radiosity approach for environments whose geometry and surface attributes.
Face Alignment with Part-Based Modeling
Robust Global Registration Natasha Gelfand Niloy Mitra Leonidas Guibas Helmut Pottmann.
Chapter 6 Feature-based alignment Advanced Computer Vision.
Model base human pose tracking. Papers Real-Time Human Pose Tracking from Range Data Simultaneous Shape and Pose Adaption of Articulated Models using.
Global spatial layout: spatial pyramid matching Spatial weighting the features Beyond bags of features: Adding spatial information.
Semi-automatic Range to Range Registration: A Feature-based Method Chao Chen & Ioannis Stamos Computer Science Department Graduate Center, Hunter College.
Reverse Engineering Niloy J. Mitra.
Registration of two scanned range images using k-d tree accelerated ICP algorithm By Xiaodong Yan Dec
Last Time Pinhole camera model, projection
Iterative closest point algorithms
Distributed Regression: an Efficient Framework for Modeling Sensor Network Data Carlos Guestrin Peter Bodik Romain Thibaux Mark Paskin Samuel Madden.
Detecting and Tracking Moving Objects for Video Surveillance Isaac Cohen and Gerard Medioni University of Southern California.
Structure from motion. Multiple-view geometry questions Scene geometry (structure): Given 2D point matches in two or more images, where are the corresponding.
 Image Search Engine Results now  Focus on GIS image registration  The Technique and its advantages  Internal working  Sample Results  Applicable.
1 3D Models and Matching representations for 3D object models particular matching techniques alignment-based systems appearance-based systems GC model.
Matching and Recognition in 3D. Moving from 2D to 3D Some things harderSome things harder – Rigid transform has 6 degrees of freedom vs. 3 – No natural.
Recovering Articulated Object Models from 3D Range Data Dragomir Anguelov Daphne Koller Hoi-Cheung Pang Praveen Srinivasan Sebastian Thrun Computer Science.
Fitting a Model to Data Reading: 15.1,
Semi-Supervised Clustering Jieping Ye Department of Computer Science and Engineering Arizona State University
Fast Isocontouring For Improved Interactivity Chandrajit L. Bajaj Valerio Pascucci Daniel R. Schikore.
Jacinto C. Nascimento, Member, IEEE, and Jorge S. Marques
TEMPLATE DESIGN © The computation of the confidence over K multiple scans is computed as if all scene points came from.
3D Models and Matching representations for 3D object models
כמה מהתעשייה? מבנה הקורס השתנה Computer vision.
3D Global Registration. The Problem Given: n scans around an objectGiven: n scans around an object Goal: align them allGoal: align them all First attempt:
Yuping Lin and Gérard Medioni.  Introduction  Method  Register UAV streams to a global reference image ▪ Consecutive UAV image registration ▪ UAV to.
Computer Vision Spring ,-685 Instructor: S. Narasimhan Wean Hall 5409 T-R 10:30am – 11:50am.
BIONFORMATIC ALGORITHMS Ryan Tinsley Brandon Lile May 9th, 2014.
Exploiting Segmentation for Robust 3D Object Matching Michael Krainin, Kurt Konolige, and Dieter Fox.
Niloy J. Mitra Leonidas J. Guibas Mark Pauly TU Vienna Stanford University ETH Zurich SIGGRAPH 2007.
Shape Matching for Model Alignment 3D Scan Matching and Registration, Part I ICCV 2005 Short Course Michael Kazhdan Johns Hopkins University.
ALIGNMENT OF 3D ARTICULATE SHAPES. Articulated registration Input: Two or more 3d point clouds (possibly with connectivity information) of an articulated.
Theory Revision Chris Murphy. The Problem Sometimes we: – Have theories for existing data that do not match new data – Do not want to repeat learning.
A Method for Registration of 3D Surfaces ICP Algorithm
MESA LAB Two papers in icfda14 Guimei Zhang MESA LAB MESA (Mechatronics, Embedded Systems and Automation) LAB School of Engineering, University of California,
Realtime 3D model construction with Microsoft Kinect and an NVIDIA Kepler laptop GPU Paul Caheny MSc in HPC 2011/2012 Project Preparation Presentation.
Applied Mathematics 1 Do we have to see everything - all the time? …or is it at all possible? Dr. ing. Rune Aasgaard.
Design of a High-Throughput Low-Power IS95 Viterbi Decoder Xun Liu Marios C. Papaefthymiou Advanced Computer Architecture Laboratory Electrical Engineering.
CS654: Digital Image Analysis Lecture 25: Hough Transform Slide credits: Guillermo Sapiro, Mubarak Shah, Derek Hoiem.
A Statistical Method for 3D Object Detection Applied to Face and Cars CVPR 2000 Henry Schneiderman and Takeo Kanade Robotics Institute, Carnegie Mellon.
EFFICIENT VARIANTS OF THE ICP ALGORITHM
A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.
Robotics Club: 5:30 this evening
Using simplified meshes for crude registration of two partially overlapping range images Mercedes R.G.Márquez Wu Shin-Ting State University of Matogrosso.
CSE 185 Introduction to Computer Vision Feature Matching.
1 CSCE 441: Computer Graphics Hidden Surface Removal Jinxiang Chai.
Improvements on Automated Registration CSc83020 Project Presentation Cecilia Chao Chen.
CSCI 631 – Foundations of Computer Vision March 15, 2016 Ashwini Imran Image Stitching.
11/25/03 3D Model Acquisition by Tracking 2D Wireframes Presenter: Jing Han Shiau M. Brown, T. Drummond and R. Cipolla Department of Engineering University.
CSCI 631 – Foundations of Computer Vision March 15, 2016 Ashwini Imran Image Stitching Link: singhashwini.mesinghashwini.me.
A Plane-Based Approach to Mondrian Stereo Matching
Semi-Supervised Clustering
CENG 789 – Digital Geometry Processing 08- Rigid-Body Alignment
Recognizing Deformable Shapes
Iterative Closest Point
The Brightness Constraint
Geometrically Stable Sampling for the ICP Algorithm
3D Models and Matching particular matching techniques
Structure from motion Input: Output: (Tomasi and Kanade)
The Brightness Constraint
3D Scan Alignment Using ICP
--- Range Image Registration
Structure from motion Input: Output: (Tomasi and Kanade)
Image Registration  Mapping of Evolution
Presentation transcript:

Automatic 3D modeling from range images Daniel Huber Carnegie Mellon University Robotics Institute

2 Outline What is 3D modeling and why automate it? Proposed approach Automatic modeling algorithms Basic modeling operations Automatic modeling example

What is 3D modeling? Applications Reverse engineering Multimedia content creation Preservation of art and architecture Terrain and environment modeling Model-based vision systems conversion to meshes pair-wise registration integration multi-view registration final model original object

4 3D modeling issues and current solutions Registration issues Which views overlap? What are the relative poses? What are the absolute poses? Mechanical solutions Measure absolute poses Measure relative poses Manual solutions Mark correspondences Hand register Specify overlapping views and verify results

5 Example application – handheld modeling User holds object to be modeled Range images are captured from various viewpoints Software determines pose relations and outputs the 3D model

6 Benefits of automatic modeling Enables new applications Handheld modeling Large-scale environment modeling Saves effort Simplifies data collection Eliminates tedious hand registration Provides an alternative to mechanical solutions Less cost & complexity Fewer limits on domain

7 Some existing modeling systems Modeling from range images Mechanical – Rioux (interiors), Miller (terrain), Levoy, Wheeler (objects) Manual – Neugebauer, Ikeuchi, Pulli (statues), Johnson (objects and interiors) Modeling from intensity images Video streams – Tomasi & Kanade, Zisserman (man- made objects) Photographs – Debevec, Zisserman (buildings), Sullivan & Ponce, Kutulakos & Seitz (objects)

8 The automatic modeling problem Don’t assume: Relative poses are known Overlapping views are known Structured scene – planes, corners, etc. Do assume: Views will overlap Views may be noisy or spurious Static scene Given a set of range images of an unstructured scene, automatically construct a consistent 3D model.

9 Overview of automatic modeling approach Select view pairs that are likely to register correctly Perform pair-wise registration and locally verify Robustly combine matches in a network of views (model graph) Infer new links based on topology Verify global consistency Improve quality by simultaneously registering all views

10 Key difficulties Don’t know what the object looks like Don’t know approximate poses or even which views overlap Pair-wise registration fails sometimes: Incorrect matches Missed matches Multiple matches (best match may be wrong) Pair-wise registration takes time – need to choose views to register Local verification cannot always detect errors – use global verification

11 Example of a locally unresolvable ambiguity overhead view views of a room pair-wise registration gives two equally good matches global consistency checks resolve the ambiguity the resulting model hypotheses

12 Surface matching process View selection Pair-wise registration Local verification Proposed automatic modeling framework Working set Pair-wise registration results Inferred matches Model construction process Topological inference Conflict detection Multi-view registration Model hypotheses selected matches Input views selected views registration results inferred matches output model

13 The model graph Encode model topology Nodes are views Arcs are transforms between overlapping views Spanning tree is the minimal complete model Perform model construction Start with no arcs Basic operations add, modify, or delete arcs Encode a priori information as initial arcs incorrect matches correct matches T 1,2

14 Automatic modeling framework Automatic modeling algorithms Model construction process Global verification Topological inference Conflict detection Multi-view registration Surface matching process View selection Pair-wise registration Local verification Basic modeling operations

15 Automatic modeling framework Automatic modeling algorithms Model construction process Global verification Topological inference Conflict detection Multi-view registration Surface matching process View selection Pair-wise registration Local verification Basic modeling operations

16 Automatic modeling algorithms Model construction difficulties One bad match corrupts the model Noisy data and spurious views Incomplete information – missed or untried matches Solutions Combinatorial optimization – deterministic and non- deterministic Study simpler problem first Exhaustive registration – try all pairs up front Selective registration – enable view selection

17 Exhaustive registration Sequential removal – start with all matches and remove Sequential addition – start with seed view and add Merging Repeatedly merge partial models Track multiple hypotheses Non-deterministic methods RANSAC – spanning tree + verification MCMC or annealing – search for maximum likelihood model

18 Selective registration View selection Use order of data collection Cluster similarly shaped views and register in clusters Sort views by shape similarity and register in order Choose self-constraining or unique views Modify exhaustive methods to include view selection

19 Automatic modeling framework Automatic modeling algorithms Model construction process Global verification Topological inference Conflict detection Multi-view registration Surface matching process View selection Pair-wise registration Local verification Basic modeling operations

20 Automatic modeling framework Automatic modeling algorithms Model construction process Global verification Topological inference Conflict detection Multi-view registration Surface matching process View selection Pair-wise registration Local verification Basic modeling operations

21 Uninformed pair-wise registration Align two views with no initial estimate Based on Johnson’s surface matching engine Uses local shape signatures – spin images Searches for similarly shaped regions Fails dramatically sometimes

22 Pair-wise registration refinement Align two views given a close initial estimate Iterative closest point algorithm (Besl & McKay, Zhang) Repeatedly minimize distance between closest points Incorrect correspondences slow convergence Minimize point to tangent plane distance instead (Chen & Medioni) point-point correspondences point-plane correspondences slow sliding fast sliding

23 Automatic modeling framework Automatic modeling algorithms Model construction process Global verification Topological inference Conflict detection Multi-view registration Surface matching process View selection Pair-wise registration Local verification Basic modeling operations

24 Local verification – overlap distance Registration requires significant overlap for success (~ 30%) Correctly registered views should be close wherever they overlap Verify matches based on overlap amount and distance Use as a classifier or as a measure of match quality Views 1 and 2: 48%, 2.1 mmViews 1 and 9: 15%, 3.1 mm

25 Local verification – visibility consistency Consider expected observations from viewpoint of sensor C 1 Consistent – S 2 close to S 1 wherever both observed Free space violation – S 2 blocks visibility of S 1 Occupied space violation – S 1 unobserved, even though it should be (sensor model required) consistent surfaces S1S1 S2S2 C1C1 C2C2 surfaces are close free space violation S1S1 S2S2 C1C1 C2C2 S 2 blocks S 1 occupied space violation S1S1 S2S2 C1C1 C2C2 S 1 is not observed

26 Model construction process Global verification Topological inference Conflict detection Multi-view registration Automatic modeling framework Automatic modeling algorithms Surface matching process View selection Pair-wise registration Local verification Basic modeling operations

27 Global verification Topological inference (positive evidence) Coarse test – bounding box intersection Fine test – local verification procedures Pair-wise register if possible/necessary Conflict detection (negative evidence) Coarse test – frustum or view volume intersection Fine test – visibility consistency

28 Automatic modeling framework Automatic modeling algorithms Model construction process Global verification Topological inference Conflict detection Multi-view registration Surface matching process View selection Pair-wise registration Local verification Basic modeling operations

29 Motivation for multi-view registration T12T12 T23T23 T34T34 T45T45 T56T56 T61T61 T 12 * T 23 * T 34 * T 45 * T 56 * T 61  I Pair-wise registration gives relative pose with small error Errors accumulate with compounding operation, leading to gaps or seams on the final model Multi-view registration distributes the accumulated error in a principled way Close-up of tail - the red, green and blue surfaces should be aligned with the cyan one top view

30 Multi-view registration Simultaneously align two or more surfaces Multi-view ICP Minimize squared distance between corresponding points (Benjemaa & Schmitt) Requires initial pair-wise estimates Suffers from same “sliding” problem as ICP Minimize point-plane distances, instead (Neugebauer)

31 Comparison of multi-view registration methods pair-wise registration only point-point correspondences point-plane correspondences Synchronized cross-section

32 Automatically modeling the angel top view connectivity after local verification horizontal slice initial model (bad matches in red)

33 Inferring new pose relations % overlap1% overlap

34 horizontal slice The final angel model top view

35 Research schedule

36 Conclusion Expected contributions Framework for automatic modeling – model graph, basic operations Algorithms for exhaustive and selective modeling Demonstration in several domains, including handheld modeling application See the proposal for Range image to mesh conversion Registration uncertainty measure Planned experiments Details of related work