Download presentation
Presentation is loading. Please wait.
Published byLindsey Kelley Modified over 9 years ago
1
Depth Estimate and Focus Recovery Presenter : Wen-Chih Hong Adviser: Jian-Jiun Ding Digital Image and Signal Processing Laboratory Graduate Institute of Communication Engineering National Taiwan University, Taipei, Taiwan, ROC 台大電信所 數位影像與訊號處理實驗室 2015/8/301DISP Lab @ MD531
2
Outlines Introduction Binocular version systems Stereo Monocular version systems DFF DFD Other method Conclusions References 2015/8/302DISP Lab @ MD531
3
Introduction Depth is an important information for robot and the 3D reconstruction. Image depth recovery is a long-term subject for other applications such as robot vision and the restorations. Most of depth recovery methods based on simply camera focus and defocus. Focus recovery can help users to understand more details for the original defocus images. 2015/8/303DISP Lab @ MD531
4
Introduction Categories of depth estimation Monocular Depth from defocus (DFD) Depth from focus (DFF) Binocular Stereo focus 2015/8/304DISP Lab @ MD531
5
Introduction Categories of depth estimation Active : Sending a controlled energy beam Detection of reflected energy Passive : Image-based 2015/8/305DISP Lab @ MD531
6
Introduction Geometric on imaging F D/2 F u s 2R : R>0 sensor v Biconvex 2015/8/306DISP Lab @ MD531
7
Binocular version systems The flow chart to binocular depth estimation. Depth map HVS modeling Edge detection Correspondence Vengeance control Gaze control Depth map 2015/8/307DISP Lab @ MD531
8
Binocular version systems Vengeance movement : is some kind of slow eye movement that two eyes move in different directions. But corresponding problem Gazing point (Corresponding point) Baseline (B) B/2 Depth (u) 2015/8/308DISP Lab @ MD531
9
Binocular version systems Complex model Corresponding point Depth (u) Right vision Left vision Baseline (B) (x R, y R ) (x L, y L ) Figure 3.3 A more complete triangulation geometry for the binocular vision. We have to realize how much departure between the optical axis and the direction of the 2015/8/309DISP Lab @ MD531
10
Binocular version systems Corresponding problem But more accuracy 2015/8/3010DISP Lab @ MD531
11
Monocular version systems Depth from focus Depth form defocus 2015/8/3011DISP Lab @ MD531
12
Depth from Focus Taking pictures at different observer distance or object distance We need an estimator to measure degree on focus Using Laplacian operator Such operator point to a measurement on a single pixel influence, a sum of Laplacian operator is needed: 2015/8/3012DISP Lab @ MD531
13
Depth from Focus Gaussian interpolation Figure 4.4 Gaussian interpolation to a measure curve, N k ≧ N k-1, N k ≧ N k+1 displacement dkdk [SML] NPNP Focus measure NkNk N k-1 d k-1 dpdp Measured curve Ideal condition d k+1 N k+1 2015/8/3013DISP Lab @ MD531
14
Depth from Focus Range from focus using Take pictures along the axis Find the image having highest frequency Need more than 10 images (monocular) 2015/8/3014DISP Lab @ MD531
15
Depth from Focus We use Gaussian interpolation to form a set of approximations The depth solution d p from above Gaussian: 2015/8/3015DISP Lab @ MD531
16
Depth from Defocus Due to geometric optics, the intensity inside the blur circle should be constant. Considering of aberration and diffraction and so on, we easily assume a blurring function: : diffusion parameter Diffusion parameter is related to blur radius: derived from triangularity in geometric optics For easy computation, we assume that foreground has equal- diffusion, background has equal-diffusion and so on However, this equal-focal assumption will be a problem 2015/8/3016DISP Lab @ MD531
17
Depth from Defocus Blurring model Blurring radius 2015/8/3017DISP Lab @ MD531
18
Depth from Defocus Blurring model 2015/8/3018DISP Lab @ MD531
19
Depth from Defocus Blurring model 2015/8/3019DISP Lab @ MD531
20
Depth from Defocus Blurring model 2015/8/3020DISP Lab @ MD531
21
Depth from Defocus Blurring model when blur radius is independent of the location of the point source on the object plane at depth 2015/8/3021DISP Lab @ MD531
22
Depth from Defocus Blurring model Using and We get So diffusion parameter: 2015/8/3022DISP Lab @ MD531
23
Depth from Defocus Depth recovery Eliminating D from m=1,2 we get where and 2015/8/3023DISP Lab @ MD531
24
Depth from Defocus Depth recovery From Take F.T.: The F.T. of Gaussian is Gaussian 2015/8/3024DISP Lab @ MD531
25
Depth from Defocus Depth recovery Take the log Using the relationship between them we get 2015/8/3025DISP Lab @ MD531
26
Depth from Defocus Depth recovery let apha=1 we obtain the value of sigma-2 Find out the depth D 2015/8/3026DISP Lab @ MD531
27
Depth from Defocus The main sources of range errors in DFD Inaccurate modeling of the optical system. Windowing for local feature analysis. Low spectral content in the scene being images. Improper calibration of camera parameters. Presence of sensor noise. 2015/8/3027DISP Lab @ MD531
28
Depth from Defocus Block shift-variant blur model Consider the interaction of sub-images Define the neighborhood function Indeed, the image we observed is compared with 2015/8/3028DISP Lab @ MD531
29
Depth from Defocus Space-variant filtering models for recovering depth Using complex spectrogram and P.W.D. Complex Spectrogram: 2015/8/3029DISP Lab @ MD531
30
Depth from Defocus Space-variant filtering models for recovering depth C.S.: g_1/g_2 where 2015/8/3030DISP Lab @ MD531
31
Depth from Defocus Space-variant filtering models for recovering depth objective function: Drawback: No consider the intersection of pixels there will be interrupt in border. Regularized solution. 2015/8/3031DISP Lab @ MD531
32
Depth from Defocus No corresponding problem Less accuracy S.V. > B.S.V. Blocking Trade-off Blocking size Too large: less accuracy Too small: noise 2015/8/3032DISP Lab @ MD531
33
Other method Structure from motion Shape from shading ML Estimation of Depth and Optimal camera settings Recursive computation of depth from multiple images 2015/8/3033DISP Lab @ MD531
34
Other method Structure from motion Using the relative motion between object and camera to find out surface information Corresponding problem (binocular) Find out what motion of camera 2015/8/3034DISP Lab @ MD531
35
Other method Shape from shading Need to know the reflectance Find the sliding rate and blindness 2015/8/3035DISP Lab @ MD531
36
Focus recovery SML measurement Defocused image pair Full focused image Maximum value searching Depth measurement of a point Small aperture construction Linear canonical transform based on constructed optical system focal point Using the specific depth to retrieve imaging distance 2015/8/3036DISP Lab @ MD531
37
Conclusions Binocular stereo method high accuracy Absolute depth information Complexity computation Corresponding problem Structure form motion Nonlinear problem Corresponding problem Shape from shading Very difficult method Active method 2015/8/3037DISP Lab @ MD531
38
Conclusions Range from focus: Slowly More than 10 images depth from defocus: Easy method Less accuracy 2015/8/3038DISP Lab @ MD531
39
References and future work 1)Y. C. Lin, Depth Estimation and Focus Recovery, Master thesis, National Taiwan Univ., Taipei, Taiwan, R.O.C, 2008 2)Subhasis Chaudhuri, A.N. Rajagopalan, ”Depth From Defocus: A Real Aperture Imaging Approach. ” Springer-Verlag. New York, 1999. 3)M. Subbarao, “Parallel depth recovery by changing camera parameters,” Second International Conference on Computer Vision 1988, pp. 149-155, Dec. 1988. 4)M. Subbarao and T. C. Wei, “Depth from defocus and rapid autofocusing: a practical approach,” IEEE Conferences on Computer Vision and Pattern Recognition, pp. 773-776, Jun. 1992. 5)A. N. Rajagopalan and S. Chaudhuri, “A variational approach to recovering depth from defocused images,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 19, pp. 1158-1164, Oct. 1997. 2015/8/3039DISP Lab @ MD531
40
The end 2015/8/30DISP Lab @ MD53140
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.