300 likes | 444 Views
Maarten Van Lier 2 e Master Computerwetenschappen. sensors in mobile devices for capturing panoramas and environment maps. Use Case. Found an awesome view View too large for one picture Take pictures with smartphone app 360° left to right / full spherical view
E N D
Maarten Van Lier 2eMaster Computerwetenschappen sensors in mobile devicesforcapturingpanoramas and environment maps
Use Case • Foundanawesome view • View toolargeforone picture • Takepictureswithsmartphoneapp • 360° left to right / full spherical view • App combines picturesinto panorama • Reasonablyfast
ProblemDescription • Makepanoramas and environment maps • Findalignmentbetween image pairs • Combine pictures
ProblemDescription (2) • Onsmartphone! • Processing power = low • Efficiency = necessary! • Interactivity • Help user takingpictures • View resultonsmartphone • Within a reasonable time
Standard approach • Takepictures • Partially overlapping • Findneighboringpictures • Findalignmentbetweenneighbors • Intensitybasedalignment • Feature basedalignment • Compositepictures
FindAlignment Pixel Based Feature Based • Findtransformation • Withlowestmisregistration • Withhighestintensity match • Minimizeerrorfunction • Search options • Full search • Hierarchical • Incremental • Detect features • Recognizablepoints • Caracterizedwith vector • SIFT, SURF, … • Match features • Findsamepoints in images • Usingdistancebetweenvectors • Findtransformation • Transforms features to corresponing features • RANSAC foroutliers • FindHomography
Smartphone Sensors • Butsmartphones have sensors! • Accelerometer, compass, gyroscope • Determineorientation of device • Useaccelerometer & compass • Useorientation! • For estimated picture location • (and forreal time “preview”)
The Overlap Approach • Takepictures • Guide user with 3D preview of estimated panorama • Save sensor data onshutter • Find overlapping regions • Usingsaved sensor data • Detect & extract features • From overlap regions • Match features • Betweencorrespondingoverlaps • Findalignment • Compositepictures
Find Overlap Regions • Usingestimated picture locations • To findneighboringpictures • To determineestimated overlap • Boundingcirclearound picture center • Bounding box around overlap region • Axisalignedvs non axisaligned
Detect & Extract Features • Onlyfrom overlapping regions of image • Large overlap • Many features => goodalignment • Expensivedetection & extraction • Small overlap • Fewer features => bad alignment • Cheaperdetection & extraction • For actualpanoramas & envmaps • Expected: large overlap regions • Actual speed gainmaynotbeverylarge
Match Features • Neighboringpictures • Usingestimatedlocations (sensor data) • Features in same overlap region • Features estimated to be close to eachother • Useestimated 3D orpolar feature locations • Butnotyetimplemented
Match Features Likely match Possiblefalse match Unlikely match Likely match
Overlap Region Test (3) • Recommendedfor most panorama apps: • About 20% oneachside=> about 70-80% whenon all sides • Butonlyanestimate=> needs to beinvestigated & tested! • Here: on average 62% overlap • Because overlap at (nearly) all sides • Expected: a 30-40% drop in time whenextracting features onlyfrom overlap regions
Results: Timings (6 pictures) Standard Approach Overlap Approach • Find features • 2383 msfor 4998 features • Matching features & calculatehomographies • 18661 ms • Total: 21044 ms • Calculate Overlap Regions • 9 msfor 12 overlaps • Find Features • 1617 msfor 3153 features • Matching features & calculatehomographies • 8252 ms • Total: 9878 => 53% less!
So… What’sNext? • Improvestitchedresult • Increase overlap region • Paper! • Second semester: • Optimizations • Port to smartphone • Check timingsonsmartphone • More testing & results • Trysomethingsimilarfor pixel basedalignment • Initiallocationusing sensors + incrementalrefinement • And maybe HDR • Writing thesis text, making poster
References • Image Alingment and Stitching: A Tutorial • (Richard Szeliski, 2006) • SURF: Speeded Up Robust Features • (HerbertBay, TineTuytelaars, Luc Van Gool, 2006) • RecognisingPanoramas • (M. Brown, D. G. Lowe, 2003) • BoofCV • http://boofcv.org