1. Francesc Moreno-NoguerComputer Vision Lab.EcolePolytechniqueFédérale de LausannePeter N. BelhumeurShree K. NayarColumbia University SIGGRAPH 2007 study Active RefocusingofImages and Videos
2. Abstract Use an activeilluminationmethod for depth estimation from a singleimage Near Far Alternate Lighting Refocused (Near) Acquired Image Computed Depth Refocused (Far)
3. Outlines Introduction Related Work Overview Projection Dot Defocus Analysis Dot Removal & Depth Estimation Realistic Refocusing Result Limits and Conclusions
4. Introduction of Refocusing Challenges of Active Refocusing Dynamic scenes Depth Estimation be done in a single frame Active illumination Full resolution depth map Projection Dot removal Partial Occlusions captured blur kernels at depth k In-focus
23. Depth Map Completion – Iterative Merging Loop: Apply Greedy alg. to group segments Merge the two most similar neighboring segments Re-computes the features of the new merged segment Iterative Merging
27. Challenge of Refocusing Partial occlusions Different parts of the lens may see different views at an object boundary Create missing region by detecting discontinuities in depth map and extending the occluded surface using texture synthesis Foreground/background transitions Pixels at depth discontinuities may receive contributions from the fr. and bg. Blend fr./bg. images within the boundary region
28. Realistic Refocusing produces better results than existing approaches Realistic Refocusing Canon + wide aperture Photoshop - blur IrisFilter Original
32. Limitations Due to Active Illumination Translucent objects exhibit subsurface scattering Blurred dots are too weak to detect Very dark Highly inclined surface (> 70°) Poor in outdoor with strong sunlight ex: the ball and the table are assigned diff. depths due to errors on segmentation errors
33. Limitations Due to sparse dots Sparsity of the depth estimation Errors in the initial segmentation of the image ex: incorrect depth due to segmentation err.
34. Conclusions Contribution Future Work An active illumination depth estimation with single Single Frame, Complete Depth Map, Texture/Textureless scenes Projected Light Patterns are Removed High resolution refocusing of images and videos Incorporate the method into digital cameras Use intra-red source for projecting the dot patter to make the depth estimation more robust in the case of highly textured scenes
本文介紹一個簡單的方法由 projector 打出去的點光源集作主動探測物體深度的工具復原後的影像 (dots removal)後 點光源探測的深度,合成細緻的 depth map再以此作 image refocusing現階段的 depth map, 已經足夠作 dynamic scenes. [Hoiem et al. 05] – automatically constructing rough scene structure from a single image
Projection dots 透過 lens, 在 scene surface 上留影. 而 camera 照的是 mirror 上的結果. same optical axis: 設計上,希望從camera 成像平面上打光出去. 有同樣的 optical axis. camera 上的成像D(dots), 可依光學原理推出 u (object 到 lens 的 depth). 另外有以下的好處Foreshortening(因透視而縮減的高度,長度) asymmetries between the camera & projector viewpoints shadows occlusions