Understanding shapes Fun with shapes

Preview:

DESCRIPTION

Understanding shapes Fun with shapes. Li Guo 2011.07.04. Exploration of Continuous Variability in Collections of 3D Shapes (Sig11) Characterizing Structural Relationships in Scenes Using Graph Kernels  (Sig11) Context-Based Search for 3D Models (SigA10) - PowerPoint PPT Presentation

Citation preview

Understanding shapesFun with shapes

Li Guo2011.07.04

• Exploration of Continuous Variability in Collections of 3D Shapes (Sig11)

• Characterizing Structural Relationships in Scenes Using Graph Kernels (Sig11)

• Context-Based Search for 3D Models (SigA10)• Shape google: Geometric words and expressions for

invariant shape retrieval (TOG11)• Making Burr Puzzles from 3D Models (Sig11)• A Geometric Study of V-style Pop-ups: Theories and

Algorithms (Sig11)• Depixelizing Pixel Art (Sig11)• Digital Micrography (Sig11)

Exploration of Continuous Variability in Collections of 3D Shapes

Authors

What(Video)

• Propose a new technique for exploring unorganized collections of 3D models

Motivation

• 3D models become more and more• Text-based search

– Many within class

• Navigating directly in descriptor space– High-dimensional– Not intuitive

• Example-based retrieval

Related work

• Morphable models and deformation modeling– Global correspondence detection remains a

challenging open problem

• Exploring shape datasets– Text keywords– Proxies– Example-based search

Selling points

• We present a template-based interface for exploring collections of similar 3D models via constrained direct manipulation.

• We introduce a novel technique to convert descriptor variability into a deformation model for a template shape without relying on correspondences between shapes.

Overview

Descriptor variability and template deformations

Shape descriptor

Shape PCA basis

DeformationSpace

PCA basisTemplatedeformation

Shape descriptor

Template selection and deformation space

• Template selection– Order the shapes by the distance to the average

descriptor– Filter the shapes have many components

• Deformation space– Template shape with C components– 6C deformation parameters(3 translation and 3

scaling)

Exploration interface

Results

Future work

• An explicit encoding of the part connectivity• A convex formulation of a similar optimization

problem• Outlier detection for shape retrieval• Analyzing the relation of discrete variability in

the shape• Extensions to our exploration interface

Characterizing Structural Relationships in Scenes Using Graph Kernels

Authors

?

What

• Represent scenes as graphs that encode models and their semantic relationships

• Applications– Finding similar scenes– Relevance feedback– Context-based model search

Motivation

• Scene comparison

Related work

• 3D Model Search• Scene Comparison

– [Harchaoui and Bach 2007] Image comparison

Spatial Relationships

Representing Scenes As Graphs

• Enclosure, Horizontal Support, Vertical Contact, Oblique Contact

Graph Comparison

• Node Kernel• Edge Kernel• Graph Kernel: [Harchaoui and Bach 2007]• Embedding the graphs in a very high

dimensional feature space and computing an inner product

Dataset

• Google 3D Warehouse– Most have scene graph– Standardize the tagging and segmentation

(mimics the method such as PASCAL,MSRC, and LabelMe [Russell et al. 2008]

Application:Relevance feedback

Find Similar Scene

Context-based model search

Comparison

Limitations

• Simple relationship• Many scenes were not reasonably segmented

Future work

• Software that is aware of the relationships expressed in 3D scenes has significant potential to augment the scene design process.

Context-Based Search for 3D Models

Authors

What

• Context search

Motivation

• 3D model search• Scene modeling• The goal of this research is to develop a

context-based 3D search engine

Related work

• Geometric Search Engines• Spatial Context in Computer Vision

– The context challenge

Dataset

• Google 3D Warehouse– Most have scene graph– Standardize the tagging and segmentation

(mimics the method such as PASCAL,MSRC, and LabelMe [Russell et al. 2008]

Overview

• Observations– All pairs of object co-occurrence across all scenes

• Spatial Relationships• Object Similarity• Model Ranking

Results

Benefit of additional supporting objects

Comparing results with and without database tags

Failure Cases

• Geometrically very similar to a relevant object but semantically very different

• Spatial relationships are overly simplistic

Future work

• Extracting more meaningful spatial relationships between objects

• Intelligently perform complex actions(意识流 )

Shape Google: geometric words and expressions for invariant shape retrieval

Authors

Alex M. Bronstein Michael M. BronsteinLEONIDAS J. GUIBAS

MAKS OVSJANIKOV

What

• Non-rigid shape search and retrieval

Motivation

• The same as before

Related work

• [Ovsjanikov et al. 2009] First introduced– Shape Google: a computer vision approach to invariant

shape retrieval.

• [Sun et al. 2009]– Feature detector and descriptor based on heat kernels

• [Behmo et al. 2008]– Taking into consideration the spatial relations between

features

• [Jain et al. 2008]– Represent shapes as compact binary codes

Feature-based methods in computer vision

• Feature detection and feature description

Overview

Results

Conclusion

• Non-rigid shape retrieval– In text retrieval methods’ spirit– Drew analogies with feature-based image

representations used in the computer vision

Making Burr Puzzles from 3D Models

Authors

? ?

What

• Burr Puzzle: 鲁班锁,孔明锁–用一种咬合的方式把木条垂直相交固定

鲁班锁

Overview

Multi-Knot Burr Puzzle

Connection types of neighboring knots

Illustrating the puzzle disassembly

Results

A Geometric Study of V-style Pop-ups: Theories and Algorithms

Authors

? ?

What (Video)

Definition

• Scaffold– A collection of planar polygons, called patches,

that are connected at straight line segments

• V-scaffold– A scaffold where each patch is labelled as either

G,B, L,R

Essential and intriguing properties of a pop-up

• The pop-up can be closed down to a flat surface and opened

up again without tearing the paper or introducing new creases

other than those in the design.

• The closing and opening of the pop-up do not need extra

forces other than holding and turning the two book pages.

• The paper does not intersect during closing or opening.

• When closed, all pieces of the pop-up are enclosed within the

book page.

Related work

• Paper crafting• Computational pop-ups

Theoretical foundation: Double-patch mechanisms

Theoretical foundation:Single-patch mechanisms (Video)

Main contribution

• A theoretical study of the geometric structure of v-style pop-ups

• Algorithmic contributions– An interactive tool for creating v-style pop-ups– An automated algorithm for constructing a v-style

pop-up from a given 3D model

Interactive design

• An interactive tool using the mechanisms discussed above

• At each step, the tool makes automated suggestions of possible locations for adding patches

Video

Automated construction

• Input: a collection of voxels• Three steps

– Patches are first constructed to cover the exterior faces of V parallel to Z axis. (S1, D2)

– Patches covering exterior faces oriented towards the positive Z axis are added. (D1)

– The ground and the backdrop of the scaffold are determined

Future work

• On the theoretical end– Improve the stability conditions– Considering the physical properties of the paper

• On the algorithmic side– Provide more intuitive popup design tools

Depixelizing Pixel Art

Authors

What

• Pixel art: digital art where the details in the image are represented at the pixel level– Video games before the mid-1990s– Icons in older desktop environments

• Convert pixel art images to a resolution-independent vector representation

Related work

• General Image Upsampling• Pixel Art Upscaling Techniques

– pixel-based and upscale the image by a fixed integer factor

• Image Vectorization

Motivation

• Conventional image upsampling and vectorization algorithms cannot handle pixel art images well

Overview

Results

Limitations

• Closer to natural images• Splines sometimes smooth certain features

too much

Future work

• In real-time manner• Improve the handling of anti-aliased input

images• Temporal upsampling of animated pixel art

images

Digital Micrography

Authors

What

Text layout (main goal)

Relaxing the alignment constraint

• Our challenge is to balance alignment with readability by selectively relaxing the alignment constraint.

Related work

• Text Art• Non-textual layout• Layout using vector fields

Selling points

• Key technical component of our work is the introduction of a novel approach for designing boundary conditions for vector fields

Algorithm Overview

Boundary Conditions Design

Results

Future work

• Expand the range of styles supported by the framework

• Text line ordering• Accelerating the method

Thank you

Recommended