senior poster

1
Image Sequences Interpolation based on Dualhomography Transformation Student:EricHuang Supervisor:Stergios Roumeliotis Background Results Conclusion Introduction: In many robotics applications, such as search, rescue or exploration, images are collected by a camera at high frame rate, e.g.,30Hz. Due to processing constraints, however, they are stored at a lower rate, e.g.,6Hz. Thus, such discontinuous image sequences are not informative enough for human to understand the explored scene. To address this problem, this project develops a program using the geometry information to stitch adjacent images and interpolate them into original image sequences. Inthis way, the generated new images sequences are able to run at a higher frequency and provide natural, smooth visual perceptions to users. Conclusions: 1. Based on thetest dataset, the new generated image sequences are able torun at original and 2x frequency and at the same time present a smooth perspective to the viewer. 2. In this project, a robust method is developed to compute the binary mask for Laplacian blending procedure, which has able to handle any overlap cases between two stitched images. 3. Estimate the subtransformation between interpolated frames using square root of homography matrix has been proved to be very accurate and has able to generate smooth transformations. 4. The program has able to check the interpolation result of each frame by computing the similarity between current frame and previous frame. Also, restart interpolation on current frame when illestimation oc c urs. 5. The program is relatively robust and successfully interpolate over 93% images in test dataset. Some sequences are dropped due to insufficient feature points, which can happen when image is too dark or too blurry. 6. The program is currently implemented based on CPU and the runtime is tolerable. About 30 sec per frame on anIntel i7, 8 GB Desktop computer. Future direction: Implement the algorithm based on GPU, which can significantly accelerate the running speed. Given sparse points cloud of the scene and camera rotation, translation matrices, explore the way to interpolate views from any viewpoint in the scene. Namely, treat the real 3D world as a black box and use only local homography transformations to represent the 2D relationship between pixels in different frames. Solve parallel effects in the interpolation procedure. Due to the translation of camerasand the naive assumption that homograph transformation ignores such depth difference, parallel effect significantly affectsinterpolation results when depth varies dramatically. Reference [1] Gao, J., Kim, S. J.,& Brown, M. S. (2011, June). Constructing image panoramas using dualhomography warping. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on (pp. 4956). IEEE. [2] Xiong, Y.,& Turkowski, K. (1998, October). Registration, calibration and blending in creating high quality panoramas. In Applications of Computer Vision, 1998. WACV'98. Proceedings., Fourth IEEE Workshop on (pp. 6974). IEEE. [3] Forsyth, D. A., & Ponce, J. (2003). A Modern Approach. Computer Vision: A Modern Approach. [4] Hartley, R., & Zisserman, A. (2003). Multiple view geometry in computer vision. Cambridge university press. [5] Chen, S. E., & Williams, L. (1993, September). View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques (pp. 279288). ACM. [6] Higham, N. J. (1987). Computing real square roots of a realmatrix. Linear Algebra and its applications, 88, 405430. Electric & Computer Engineering College of Science & Engineering Approach Dualhomography Transformation Interpolation New image sequences that can run at original or even higher frequency Original discontinuous image sequences Stitching Homography Matrix related two images For eachpair of consecutive images, the views are not strictly related by a homography transformation due to the translation between their viewpoints. The 3D structure of scene affectsthis transformation. However, we can approximate this by using several different homography transformations in spatial layout of images. Real Sc hur dec omposition In order to Interpolate the the transformation, we decompose homography matrix by computing its real square root. f(A) function computes its square root, where Q is real orthogonal matrix.And T inherits R’s upper quasitriangular structure. Kmeans In order to divide the image in special layout, we apply 3demissional kmeanson the scaled 3d coordinates [x’, y’ , z’] for eachfeature point extracted from image pairs. The relative depth Z of each feature point is computed using optical flow. Then, for each cluster of feature points, compute local homography matrix using RANSAC and leastsquare method. Algorithm: Kmeans cluster result Cluster feature points based on kmeans with 2 bins Dualhomgraphy transformation result Stitched images without blending L aplac ian Blending result Blend images and eliminate seamline

Transcript of senior poster

Page 1: senior poster

Image  Sequences  Interpolation  based  on  Dual-­‐homographyTransformation

Student:EricHuangSupervisor:Stergios Roumeliotis

Background Results ConclusionIntroduction:In  many   robotics  applications,   such  as   search,   rescue  or  exploration,   images  are  collected  by  a  camera  at  high   frame  rate,  e.g.,  30Hz.  Due   to  processing   constraints,  however,   they  are  stored  at  a  lower   rate,  e.g.,  6Hz.  Thus,   such  discontinuous   image  sequences   are  not   informative  enough   for  human  to  understand   the  explored   scene.  

To  address   this   problem,   this   project  develops   a  program  using   the  geometry  information   to  stitch  adjacent   images   and  interpolate   them   into  original   image  sequences.   In  this  way,   the  generated  new   images  sequences   are  able   to   run  at  a  higher   frequency  and  provide   natural,   smooth   visual  perceptions   to  users.

Conclusions:1. Based on the test dataset, the new generated image sequences are able to run at

original and 2x frequency and at the same time present a smooth perspective tothe viewer.

2. In this project, a robust method is developed to compute the binary mask forLaplacian blending procedure, which has able to handle any overlap casesbetween two stitched images.

3. Estimate the sub-­‐transformation between interpolated frames using square rootof homography matrix has been proved to be very accurate and has able togenerate smooth transformations.

4. The program has able to check the interpolation result of each frame bycomputing the similarity between current frame and previous frame. Also, restartinterpolation on current frame when ill-­‐estimation occurs.

5. The program is relatively robust and successfully interpolate over 93% images intest dataset. Some sequences are dropped due to insufficient feature points,which can happen when image is too dark or too blurry.

6. The program is currently implemented based on CPU and the runtime is tolerable.About 30 sec per frame on an Intel i7, 8 GB Desktop computer.

Future direction:• Implement the algorithm based on GPU, which can significantly accelerate therunning speed.•Given sparse points cloud of the scene and camera rotation, translation matrices,explore the way to interpolate views from any viewpoint in the scene. Namely, treatthe real 3D world as a black box anduse only local homography transformations torepresent the 2D relationship between pixels in different frames.• Solve parallel effects in the interpolation procedure. Due to the translation ofcamerasand the naive assumption that homograph transformation ignores suchdepth difference, parallel effect significantly affects interpolation results when depthvaries dramatically.

Reference[1] Gao,   J. ,  Kim,   S.  J. ,  &  Brown,   M.  S.  (2011,   June).  Constructing   image  panoramas  using   dual-­‐homography warping.   In  Computer  Vision   and  Pattern  Recognition   (CVPR),  2011   IEEE  Conference   on (pp.  49-­‐56).   IEEE.

[2] Xiong,   Y.,  &  Turkowski,   K.  (1998,   October).  Registration,   calibration   and  blending   in  creating  high  quality   panoramas.   In  Applications   of  Computer  Vision,   1998.  WACV'98.  Proceedings.,   Fourth   IEEE  Workshop   on (pp.  69-­‐74).   IEEE.

[3] Forsyth,   D.  A.,  &  Ponce,   J.  (2003).  A  Modern   Approach.   Computer  Vision:   A  Modern  Approach.

[4] Hartley,  R.,  &  Zisserman,   A.  (2003).  Multiple   view  geometry   in  computer  vision.  Cambridge   university   press.

[5] Chen,   S.  E.,  &  Williams,   L.  (1993,  September).  View   interpolation   for   image  synthesis.   In  Proceedings   of   the  20th  annual   conference   on  Computer  graphics   and  interactive   techniques (pp.  279-­‐288).  ACM.

[6] Higham,   N.  J.  (1987).  Computing   real   square   roots   of  a   real  matrix.  Linear  Algebra  and   its  applications,   88,  405-­‐430.

Electric & Computer EngineeringCollege of Science &

Engineering

ApproachDual-­‐homography Transformation

Interpolation

New image   sequences that canrun at original or even higher

frequency

Original discontinuous imagesequences

Stitching

Homography Matrix relatedtwo images

For eachpair of consecutive images, the viewsare not strictly related by a homographytransformation due to the translation betweentheir viewpoints. The 3D structure of sceneaffects this transformation. However, we canapproximate this by using several differenthomography transformations in spatial layout ofimages.

Real   Schur decompositionIn order to Interpolate the the transformation,we decompose homography matrix bycomputing its real square root.f(A) function computes its square root, where Qis real  orthogonal   matrix.And T  inherits   R’s  upper  quasitriangular structure.

K-­‐meansIn order todivide the image in special layout, weapply 3-­‐demissional k-­‐meanson the scaled 3dcoordinates [x’, y’, z’] for each feature pointextracted from image pairs.The relative depth Z of each feature point iscomputed using optical flow.Then, for each cluster of feature points, computelocal homography matrix using RANSAC andleast-­‐square method.

Algorithm:

K-­‐means cluster result

Cluster   feature  points  based  on  k-­‐means  with   2  bins

Dual-­‐homgraphytransformation result

Stitched imageswithout blending

Laplacian Blending result

Blend images andeliminate seamline