Hauptseminar Bildanalyse SS 16

Dieses Semester:

Meine 3D Urlaubsfotos im 3D Fernseher - und die Bildverarbeitung die dafür nötig sein kann.

Angeboten in den Modulen:

INF-AQUA, INF-BAS2, INF-BAS7, INF-D-940, INF-VERT2, INF-VERT7, IST-05-HS, MINF-04-HS

Sommersemester 2016

Termin

Donnerstag, 2.DS, 9:20 - 10:50 Uhr, ABP 2026

Beginn: Donnerstag, 07. April 2016

Leitung

Prof. Carsten Rother, Holger Heidrich

Beschreibung

Im Seminar werden aktuelle Themen der Computer Vision Literatur behandelt, die im Arbeitsgebiet unsere Gruppe liegen. Es ist ein Vortrag von 30min zu halten und eine schriftliche Arbeit in der Form eines Reviews zu verfassen.
Der Zeitplan im Seminar ist:
  • Problemvorstellung (5 min.)
  • Diskussion (10min.)
  • Vorstellung der Lösung/Methodik des Artikels (25min.)
  • Discussion (15min.)

  • Abgabe: Vortragsfolien und Review.
  • Bewertung: Vortrag (60%), Review (20%), Diskussion im Seminar (20%)
Ziele: Es ist eine Konsultation beim Betreuer zu vereinbaren, die spätestens eine Woche vor dem Vortrag liegen muss.
Not mandatory - but if you can please give your talk in English.

Einschreibung:

Einschreibung über jExam, Themenwahl über TUD-email an Holger.Heidrich
Terminvergabe zum ersten Seminar entsprechend der Einschreibreihenfolge.

Inhalt: 3d Scene Reconstruction und Image Matching

Themenangebot:
  1. Signal and noise in single and stereo images
      For stereo we like to match points in the two images. We do this on the basis of local intenisty distributions (features). But we need to distinguish these from noise. Therefore we need to know what noise we have.
    1. Noise Estimation from a Single Image, longer version, ppt.
    2. Single-image Noise Level Estimation for Blind Denoising, Matlab code.
    3. A model for measurement of noise in CCD digital-video cameras, Bearbeiterin: Zheng Jing, Termin: 02.06.2016
    4. Joint Noise Level Estimation from Personal Photo Collections.
    5. EMVA 1288: Standard for Measurement and Presentation of Specifications for Machine Vision Sensors and Cameras.
    6. Is Denoising Dead?.
  2. Image function estimation from noisy and sampled images (single and stereo)
      If we know (an estimation of) the noise we want to recover the image function as good as possible. (First 2 papers are the same as for noise estimation as these also try to remove the noise.)
    1. Noise Estimation from a Single Image, longer version, ppt.
    2. Single-image Noise Level Estimation for Blind Denoising, Matlab code, Bearbeiter: Michael Jobst, Termin: 09.06.2016.
    3. Adaptive-neighborhood filtering of images corrupted by signal-dependent noise .
    4. Image Denoising Methods. A New Nonlocal Principle (see also newer papers on that side), test and code: Parameter-Free Fast Pixelwise Non-Local Means Denoising
    5. Bayesian Deblurring with Integrated Noise Estimation
    6. The Noise Clinic: a Blind Image Denoising Algorithm
    7. Gaussian Process Random Fields (a method to get an interpolated function from noisy data).
  3. Segmentation, grouping, cartoonization, edge aware filtering and edge preserving pyramids for single and stereo images
      It would be helpful to know the objects in an image. These cause depth jumps and occlusions. Generally we assume that homogenously textured regions belong to one and the same object. Still we have to specify what homogenously textured means (think of a picture of trees e.g.).
    1. Displets: Resolving Stereo Ambiguities using Object Knowledge , Bearbeiter: Lucas Kahlert, Termin: 12.05.2016
    2. Deep Filter Banks for Texture Recognition, Description, and Segmentation
    3. Superpixel Meshes for Fast Edge-Preserving Surface Reconstruction
    4. Efficient Joint Segmentation, Occlusion Labeling, Stereo and Flow Estimation
    5. Viewpoint Invariant Texture Matching andWide Baseline Stereo
    6. Stereo matching with superpixels , Bearbeiter: Michael Engel, Termin: 26.05.2016
    7. Superpixels, Occlusion and Stereo
  4. Blur aware matching
      There is some "natural blur" because image acquisition (hopefully) followed the sampling theorem and there may be additional blur due to defocus (varying depth) or motion. Yet we would like to have subpixel accurate disparities.
    1. Blur-aware Disparity Estimation from Defocus Stereo Images, Bearbeiterin: Franziska Krüger, Termin: 16.06.2016
    2. Does Color Really Help in Dense Stereo Matching?
    3. Real-time Local Stereo Matching Using Guided Image Filtering , Bearbeiter: Kurt Lachmann, Termin: 23.06.2016
  5. Features and subpixel accurate correspondence detection for stereo images
      Depth, and therefore disparities are continuous variables. Integer disparities may be easier to calculate, but generally subpixel disparities are possible (whenever integer disparities can be deduced). Most accurate subpixel disparity estimation is possible in regions with constant disparity and a lot of structure, i.e. strong high frequency components (up to the Nyquist frequency). But this is only given in high textured fronto-parallel planes. Because these are rare we either stick to very local structures (lines, edges) or allow transforms between local patches. A commonly used transform is a homography between local plane patches. To find such regions, feature detectors (and descriptors) are used.
    1. A Local Algorithm for the Computation of Image Velocity via Constructive Interference of Global Fourier Components,Babette Dellen, Florentin Wörgötter, 2011.,
              Bearbeiter: Robert Wünsche, Termin: 30.06.2016
    2. The Design and Use of Steerable Filters
    3. The Naked Truth about Cost Functions for Stereo Matching
    4. Reliability and accuracy in stereovision Application to aerial and satellite high resolution images
  6. Vanishing points for stereo correspondences
      Vanishing points are the same in rectified stereo images. This can be used to support correspondence detection between straight lines.
    1. A Minimal Closed-form Solution for the Perspective Three Orthogonal Angles (P3oA) Problem: Application To Visual Odometry
  7. View dependend matching
  8. robust Model Estimation
    1. Tractable Algorithms for Robust Model Estimation, Enqvist, Olof and Ask, Erik and Kahl, Fredrik and Åström, Kalle, (2014).
  9. (data driven) sequential optimal approaches
  10. fast nearest neighbour access: {hashing, subspace methods, ...}
  11. more to come
Seminarplan Sommersemester 2016

THEMA TERMIN VORTRAGENDE, VORTRAGENDER
Einführung 07.04.2016 Holger Heidrich
Displets: Resolving Stereo Ambiguities using Object Knowledge 12.05.2016 Lucas Kahlert, Vortrag
Stereo matching with superpixels 26.05.2016 Michael Engel
A model for measurement of noise in CCD digital-video cameras 02.06.2016 Zheng Jing
Single-image Noise Level Estimation for Blind Denoising 09.06.2016 Michael Jobst, Vortrag
Blur-aware Disparity Estimation from Defocus Stereo Images 16.06.2016 Franziska Krüger, Vortrag
Real-time Local Stereo Matching Using Guided Image Filtering 23.06.2016 Kurt Lachmann, Vortrag
A Local Algorithm for the Computation of Image Velocity via Constructive Interference of Global Fourier Components 30.06.2016 Robert Wünsche