Radiometric Self Calibration Tomoo Mitsunaga Shree K. Nayar Hashimoto Signal Processing Lab. Sony Corporation Dept.
Download ReportTranscript Radiometric Self Calibration Tomoo Mitsunaga Shree K. Nayar Hashimoto Signal Processing Lab. Sony Corporation Dept.
Radiometric Self Calibration Tomoo Mitsunaga Shree K. Nayar Hashimoto Signal Processing Lab. Sony Corporation Dept. of Computer Science Columbia University CVPR Conference Ft. Collins, Colorado June 1999 Problem Statement How well does the image represent the real world? Image M2 (Low exposure) Image M1 (High exposure) Usual imaging systems have : • Limited dynamic range • Non-linear response ( M 2 M1 ) June/1999 CVPR99 2 Scene Radiance and Image Irradiance E L Radiance Irradiance d 2 Image irradiance : d E L cos kL 4h 2 Ideal camera response : I E t Exposure : 2 4 Aperture area 2 d e t 2 I kL e June/1999 CVPR99 3 Scene Radiance and Measured Brightness Video Image Formation Scene radiance L Image Exposure linear CCD Scaled radiance I Camera Electronics Digitization I f (M ) M f 1 ( I ) Measured brightness M Photo Image Formation Image Exposure Film Film Development Scanning f (M) : The radiometric response function June/1999 CVPR99 4 Calibration with Reference Objects The scene must be controlled • The reflectance of the objects must be known • The illumination must be controlled June/1999 CVPR99 5 Calibration without Reference Objects • Differently exposed images from an arbitrary scene • Recover the response function from the images • Calibrate the images with the response function 1 0.8 Input Images 0.6 Response function 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 High dynamic range radiance image June/1999 CVPR99 6 Previous Works • Mann and Picard (95) : – Take two images with known exposure ratio R – Restrictive model for f : M a bg I – Find parameters a, b, g by regression • Debevec and Malik (97) : – General model for f : only smoothness constraint – Take several (say, 10) high quality images – At precisely measured exposures (shutter speed) June/1999 CVPR99 7 Obtaining Exposure Information We have only rough estimates • Mechanical error • Reading error (ex. F-stop number) June/1999 CVPR99 8 Radiometric Self-Calibration Works with roughly estimated exposures • Inputs : – Differently exposed images – Rough estimates of exposure values – ex. F-stop reading • Outputs : – Estimated response function – Corrected exposure values June/1999 CVPR99 9 A Flexible Parametric Model 1 High order polynomial model : 0.8 video N I f ( M ) cn M n 0.6 f (M) n 0 s 0.4 posi 0.2 Parameters to be recovered : nega • Coefficients cn • Order N 0 0 0.2 0.4 0.6 0.8 1 M f(M) of some popular imaging products June/1999 CVPR99 10 Response Function and Exposure Ratio Images: q = 1,2,….Q , Pixels: p = 1, 2, …..P Exposure ratio: Rq ,q 1 eq eq 1 , Rq ,q 1 L p k p eq L p k p eq 1 I p ,q I p ,q 1 n0 cn M p,q N Using polynomial model : N f ( M p ,q ) f ( M p ,q 1 ) n c M p ,q 1 n Rq ,q 1 n 0 n Thus, we obtain ... Q 1 P Objective function : June/1999 n n cn M p ,q Rq ,q 1 cn M p ,q 1 q 1 p 1 n 0 n 0 N CVPR99 N 2 11 An Iterative Scheme for Optimization Rough estimates Rq,q+1(0) Rq,q+1(i) Optimize for Rq,q+1 Optimize for f f (i) f (i ) f (i 1) M Optimized f and Rq,q+1 June/1999 CVPR99 12 Evaluation : Noisy Synthetic Images 1 14 16 0.8 25 34 48 51 64 0.6 72 91 99 f (M) 14 16 0.4 25 34 48 51 64 0.2 72 91 99 0 0 0.2 0.4 0.6 0.8 1 M Solid : Computed response function Dots : Actual response function June/1999 CVPR99 13 Evaluation : Noisy Synthetic Images (cont’d) 10 8 Percentage Error in Computed Response Function 6 4 2 0 0 10 20 30 40 50 60 70 80 90 100 Trial Number Maximum Error : 2.7 % June/1999 CVPR99 14 Computing a High Dynamic Range Image • Calibrating by the response function f (M ) • Normalizing by corrected exposure values e~ • Averaging with SNR-based weighting w(M ) M p ,1 M p,2 f (M ) f (M ) I p ,1 I p,2 1~ e1 ~ I p ,1 1~ e2 ~ I p,2 ~ w ( M ) I q1 p,q p,q Q Q ~ I p ,Q M p ,Q June/1999 f (M ) I p ,Q q 1 Ip w( M p ,q ) 1~ eQ CVPR99 15 Results : Low Library (video) Captured images 1 0.8 0.6 I 0.4 0.2 Calibration chart 0 0 0.2 0.4 0.6 0.8 1 M Computed response function June/1999 CVPR99 16 Results : Low Library (video) Captured images Computed radiance image June/1999 CVPR99 17 Results : Adobe Room (photograph) Captured images 1 0.8 0.6 I 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 M Computed response function June/1999 Computed radiance image CVPR99 18 Results : Taos Clay Oven (photograph) Captured images 1 0.8 0.6 I 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 M Computed response function June/1999 Computed radiance image CVPR99 19 Conclusions A Practical Radiometric Self-calibration Method • Works with – Arbitrary still scene – Rough estimates of exposure • Recovers – Response function of the imaging system – High dynamic range image of the scene Software and Demo http://www.cs.columbia.edu/CAVE/ June/1999 CVPR99 20