1 / 19

3-D Depth Reconstruction from a Single Still Image

3-D Depth Reconstruction from a Single Still Image. 何開暘 2010.6.11. Visual Cues for Depth Perception. Monocular Cues Texture variations, texture gradients, interposition, occlusion, known object sizes, light and shading, haze, defocus Stereo Cues Motion Parallax and Focus Cues.

locke
Télécharger la présentation

3-D Depth Reconstruction from a Single Still Image

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. 3-D Depth Reconstruction from a Single Still Image 何開暘 2010.6.11

  2. Visual Cues for Depth Perception • Monocular Cues • Texture variations, texture gradients, interposition, occlusion, known object sizes, light and shading, haze, defocus • Stereo Cues • Motion Parallax and Focus Cues

  3. image → feature → depth • Chose features that capture 3 types of cues: texture variations, texture gradients, and color • Model conditional distribution of depths given monocular image features p(d|x) • Estimate parameters by maximizing conditional log likelihood of training data • Given an image, find MAP estimate of depths

  4. Outline • Introduction • Feature Vector • Probabilistic Model • Experiments • Reference

  5. Feature vectors • Two types of features • Absolute depth features―used to estimate absolute depth at a particular patch • Relative features―used to estimate relative depths • Capture three types of cues • Texture variation―apply Law’s masks to intensity channel • Haze―apply a local averaging filter to color channels • Texture gradient―apply six oriented edge filters to intensity channel

  6. Features for Absolute Depth • Compute summary statistics of a patch i in the image I(x,y) as follows • Use the output of each of the 17 (9 Law’s masks, 2 color channels and 6 texture gradients) filters Fn, n=1,…,17 as: (dimension 34) • To estimate absolute depth at a patch, local image features centered on the patch are insufficient • Use more global properties

  7. More Global Properties • Use image features extracted at multiple spatial scales (three scale) • Features used to predict depth of a particular patch are computed from that patch as well as 4 neighboring patches (Repeated at each of the three scales) • Add to features of a patch additional summary features of the column it lies in (5*3+4)*34=636 dimensional

  8. Features for Relative Depth • To learn the dependencies between two neighboring patches • Compute a 10-bin histogram of each of the 17 filter outputs , giving a total of 170 features yis for each patch i at scale s • Relative depth features yijs for two neighboring patches i and j at scale s will be the differences between their histogram, i.e., yijs=yis-yjs

  9. Outline • Introduction • Feature Vector • Probabilistic Model • Experiments • Reference

  10. Gaussian Model

  11. Laplacian Model

  12. Outline • Introduction • Feature Vector • Probabilistic Model • Experiments • Reference

  13. Result

  14. Improving Performance of Stereovision using Monocular Cues

  15. The average errors as a function of the distance from the camera

  16. Reference • A.Y. Ng A. Saxena, S.H. Chung. 3-d depth reconstruction from a single still image. In International Journal of Computer Vision (IJCV), 2007. • Michels, J., Saxena, A., & Ng, A. Y. (2005). High speed obstacle avoidance using monocular vision and reinforcement learning. In 22nd international conference on machine learning (ICML).

More Related