Query Interfaces for a 3D Model Search Engine Patrick Min Tom Funkhouser, Misha Kazhdan, Joyce Chen, Alex Halderman, David Dobkin, David Jacobs DLI2 meeting, Edinburgh, 25 June 2002 Online Data On-line multimedia data is changing the way we get and use information Registered Saddlebred out of Famous Sultan Supreme line. 100% sound. 16 year old, flashy, chestnut w/white, loving, high energy horse, needs experienced rider. Was shown professionally in early years as gaited saddlebred. Most recently shown and always placed in training and first level dressage shows.Currently used as dressage/pleasure horse, jumps, loves trailriding. Text www.dreamhorse.com 2D Image What about 3D Models? Audio 3D Models are becoming easier to create, use and distribute 3D Models are becoming easier to create, use and distribute • Laser range scanners: supply Cyberware Scanner 3D Models are becoming easier to create, use and distribute • Laser range scanners: supply • Fast graphics cards: demand • World Wide Web: distribution Georgia Tech Cyberware Scanner Protein Data Bank Large Model Archive 3D Models are becoming easier to create, use and distribute • Laser range scanners: supply • Fast graphics cards: demand • World Wide Web: distribution One day 3D models will be as ubiquitous as images are now Georgia Tech Cyberware Scanner Protein Data Bank Large Model Archive Research Challenges Current research: • How do we construct 3D models ? Digital Michelangelo Project Stanford Graphics Laboratory Research Challenges Current research: • How do we construct 3D models ? Digital Michelangelo Project Stanford Graphics Laboratory Future research: • How do we find 3D models ? • How do we understand them ? Finding 3D Models car Project Goal A Search Engine for 3D Models on the Web car 3D Model Search Engine World Wide Web Crawler Crawler Crawler Crawler Repository of 3D Models Off-line On-line Indexer Text Index 2D Index 3D Index Text Matcher 2D Matcher 3D Matcher Query User Query Interface Matches 3D Model Search Engine Acquisition World Wide Web Crawler Crawler Crawler Crawler Repository of 3D Models Off-line On-line Indexer Text Index 2D Index 3D Index Text Matcher 2D Matcher 3D Matcher Query User Query Interface Matches 3D Model Search Engine Analysis World Wide Web Crawler Crawler Crawler Crawler Repository of 3D Models Off-line On-line Indexer Text Index 2D Index 3D Index Text Matcher 2D Matcher 3D Matcher Query User Query Interface Matches 3D Model Search Engine World Wide Web Crawler Crawler Crawler Crawler Repository of 3D Models Off-line On-line Indexer Text Index 2D Index 3D Index Text Matcher 2D Matcher 3D Matcher Query User Query Interface Matches Matching 3D Model Search Engine World Wide Web Crawler Crawler Crawler Crawler Repository of 3D Models Off-line On-line Indexer Text Index 2D Index 3D Index Text Matcher 2D Matcher 3D Matcher Query User Query Interface Querying Matches 3D Model Search Engine Analysis World Wide Web Crawler Crawler Crawler Crawler Repository of 3D Models Off-line On-line Indexer Text Index 2D Index 3D Index Text Matcher 2D Matcher 3D Matcher Query User Query Interface Querying Matches Matching Main Challenge & today's topic Which Query Interfaces are Useful for a 3D Model Search Engine? Query User Query Interface Matches Desired Properties of Query Interfaces easy to use • also for novice users possible to quickly specify query efficient matching method effective • high number of relevant results Available Data for Queries Example of a 3D Model available on the Web: http://www.3dspline.com Available Data for Queries Text Appearance Shape Available Data for Queries Text Appearance Shape Text Queries Text Queries + easy to use + possible to quickly specify query + efficient matching method +/- effective Effective Text Queries 747 caffeine dog dna fork Text Query Problems objects are badly annotated: • ferrar03.3ds, ferrair, rahmen, … keywords are too common • face, plane Text Query Problems objects are badly annotated: • ferrar03.3ds, ferrair, rahmen, … keywords are too common • face, plane keywords are unknown to user, or annotator not possible to anticipate every relevant text query • elf, fairy-tale creature, puppet, person, ... Text Query Problems objects are badly annotated: • ferrar03.3ds, ferrair, rahmen, … keywords are too common • face, plane keywords are unknown to user, or annotator not possible to anticipate every relevant text query • elf, fairy-tale creature, puppet, person, ... We hypothesize shape can help Outline Introduction Text Queries 3D Shape Queries 2D Shape Queries Results Conclusions Future Work 3D Shape Queries use 3D shape to find 3D shape 3D Model Queries (create &) upload 3D model select 3D model on results page ("shape iteration") 3D Matching Method 3D Matching Method Shape Iteration User Study Shape Iteration User Study How does shape combine with text in interactive searches? Test Design: • 18 students in an introductory CS class • input: on website, initial text keyword search, then “find similar object” • where similar means similar text or similar shape (unknown to user) Shape Iteration Results Similarity shape iteration finds target object more often in fewer iterations and in less time 3D Model Query Problems upload 3D model: • such a model may not be available • creating one with standard tools is hard shape iteration: • have to get such a result first 3D Sketch Queries Sketch 3D model using simple modeling program • e.g. Teddy, Sketch Teddy: A Sketching Interface for 3D Freeform Design Igarashi et al., SIGGRAPH 1999 3D Sketch Queries Teddy: A Sketching Interface for 3D Freeform Design Igarashi et al., SIGGRAPH 1999 Teddy Sample Results 3D Sketch Query Problems Teddy: supports only coarse, blobby shapes of genus 0 still hard to use for novice users 3D Sketch Query Properties - easy to use - possible to quickly specify query + efficient matching method +/- effective So we need a simpler query interface 2D Sketch Queries 2D Sketch Queries “paintbrush” style interface: • simple outline sketches • easy to use for most users Multiple 2D Sketches 2D Sketch Matching Method 5 4 6 3 7 1 Query sketch 2 7 projections per database model 2D Sketch Matching Method compute 2D shape descriptor in similar fashion intersect with circles instead of spheres 2D Sketch User Study 2D Sketch Interface User Study Does shape add to text? Test design: • 43 students in introductory CS class • goal: retrieve target objects • input: text keywords and sketches on paper Example Input 2D Sketch User Study Results Distinguishing keywords: • text works well • even better with sketch 2D Sketch User Study Results Class is too large: • text is too inspecific • sketch works best 2D Sketch User Study Results text and shape can complement each other sketch allows refinement within a class 2D Sketch Query Problems The drawing is free-form: • hard to draw accurately • hard to get proportions right Dog DC10 2D Sketch Query Problems The drawing is free-form: • hard to draw accurately • hard to get proportions right Dog DC10 Image matching method: • user and database outline have to look similar Search Engine Usage Results Search Engine Results online since early November 2001 20,707 models indexed, 2,873 commercial one typical week of usage: • 4,522 queries • 1,346 unique hosts from 55 different countries • 1,029 models downloaded Search Engine Results Text vs. Shape text shape empty about 70 % text, 30 % shape-based Search Engine Results Shape Query Type 2D text & 2D 3D text & 3D similar shape upload model about 58 % shape iteration, 38 % 2D/text & 2D Number of Hosts per Day jan feb mar apr may average about 175 hosts and 525 searches/day % New Hosts per Day jan feb mar apr about 20 % returning hosts per day may # Searches per User # hosts 7000 6000 5000 4000 3000 2000 1000 1 3 5 7 9 11 13 15 17 19 21 23 25 # searches about 40 % does 3 or more searches per day Conclusions and Future Work Conclusions shape-based search can help each interface has its strengths and weaknesses • so useful to combine them table find similar shape Future Work 2D Structural Interface primitive drawing user study evaluation of several interfaces • oval + rectangle • skeletal graphs 3D Structural Interface Future Work II 3D Model Database fully automated acquisition and model processing support for more model formats Acknowledgements Funding NSF (CCR-0093343, 11S-0121446, CCR-99-88173, DGE-9972930) Army Research Organization (DAAD19-99-1-0205) Alfred P. Sloan Fellowship (for Prof. Tom Funkhouser) People Brian Kernighan and the students of CS109 and CS111 Bernard Chazelle, Adam Finkelstein, Szymon Rusinkiewicz, Emil Praun Princeton CS department technical staff Princeton Shape Analysis Group 3D Models Viewpoint Datalabs, http://www.viewpoint.com De Espona Infografica, http://www.deespona.com http://shape.cs.princeton.edu 3D Model Search Engine Test Database 50 models, 10 categories of 5 models each