Computer Vision 2 (Summer Term 2020)

Overview

• Online course consisting of self-study and exercises (2/2/0). No audio/video lectures.
• Modules: CMS-CLS-ELG, CMS-VC-ELV1, CMS-VC-ELV2, INF-04-FG-IS, INF-B-520, INF-B-540, INF-BAS-2, INF-BAS-7, INF-E-3, INF-LE-MA, INF-PM-ANW, INF-PM-FOR, INF-VERT-2, INF-VERT-7, INF-VMI-8, MATH-MA-INFGDV
• Lecturer: Bjoern Andres
• Assistant: Holger Heidrich
• Online registration is possible via SELMA (CMS), jExam or OPAL
• Forum and exam scheduling (OPAL)

Lecture

Material for self-study is made available here as the course progresses.

• Bilateral filtering
• Pixel labeling
  • Pixel classification
  • Prior knowledge
    • Smoothness
• Image decomposition
• Joint image decomposition and pixel labeling
• Object recognition

Examinations

• Format
  • Individual remote oral examinations about the contents of this course are offered for all participating students of all degree programmes.
  • There will not be any written examination on the contents of this course in the Summer Term 2020.
• Creditability
  • Regulations regarding the creditability of the results of the examination depend on the degree programme and module. Regulations we are aware of will be summarized below. Students in doubt are asked to check with the coordinator of their degree programme.
  • CMS-*. In the degree programme Computational Modeling and Simulation (CMS), the individual remote oral examination replaces the written examination.
  • INF-BAS-*, INF-E-*, INF-PM-ANW, INF-PM-FOR, INF-VERT-*, INF-VMI-*. In these modules, there is an additional, exceptional opportunity in the Summer Term 2020 to take separate examinations on the contents of each course creditable toward the module, instead of taking one complex examination on the contents of multiple courses. A prerequisite, however, is that all these examinations are taken this term. In the future, remote oral examinations on the contents of multiple courses, possibly involving multiple examiners, can be scheduled individually as usual, by email, as described here.
• Scheduling open from 2020-05-28 through 2020-06-28
  • Appointments for individual remote oral examinations are available now and can be scheduled exclusively through OPAL.
• Registration open through 2020-07-03 (all forms need to be submitted by then)
  • Only scheduled examinations can be registered.
  • Students enrolled in the degree programme Computational Modeling and Simulation (CMS) need to
    • register their scheduled examination via SELMA
    • send the filled in and signed form Declaration of consent to an alternative form of an immaterial/oral examination to mlcv-exams@tu-dresden.de.
  • Students enrolled in other degree programmes of the Faculty of Computer Science
    • who wish to credit the course toward one of the modules INF-BAS-*, INF-VERT-*, INF-VMI-* or INF-E-* need to
      • follow these instructions by the examination office of the faculty until 2020-07-03
      • send the filled in and signed form Declaration of consent to an alternative form of an immaterial/oral examination to mlcv-exams@tu-dresden.de.
    • who wish to credit the course toward any other module need to send to mlcv-exams@tu-dresden.de:
      • The filled in and signed registration form of the module toward which the course is to be credited. Respective forms are provided by the Examination Office and by the Service Center for International Students
      • The filled in and signed form Declaration of consent to an alternative form of an immaterial/oral examination
  • Students enrolled in degree programmes of other faculties of TU Dresden need to email to mlcv-exams@tu-dresden.de:
    • Registration documents issued by the faculty of their degree programme, in case such documents exist
    • In case such documents do not exist: An informal request of a course certificate, including full name, matriculation number, degree programme, course title and email address to be used for correspondence
    • The filled in and signed form Declaration of consent to an alternative form of an immaterial/oral examination.

Preparation

The following questions are similar in style to questions that are asked during the examination. They might be of help when preparing for the examination. We emphasize, however, that the examination is in no way constrained by, let alone limited to these questions.

• Bilateral filter
  • What do you know about the bilateral filter?
  • Which metrics/weight functions appear in the definition of the bilateral filter?
  • What is their effect?
  • Describe the computations involved in the bilateral filtering of a gray scale image!
  • How does the number of operations change if we double the size of the image/filter mask?
• Pixel classification
  • What do you know about the pixel classification problem?
  • Suppose we know a priori that neighboring pixels are more likely to be of the same class. How do we incorporate this prior knowledge into the pixel classification problem?
  • Describe the objective function of the smooth pixel classification problem!
  • How do you solve the smooth pixel classification problem?
  • How do you reduce the smooth pixel classification problem to a minimum st-cut problem?
  • How do you solve the minimum st-cut problem?
• Image decomposition
  • What do you know about the image decomposition problem?
  • How does a feasible solution to the image decomposition problem look like?
  • How is image decomposition different from pixel classification?
  • What is a decomposition/multicut of an image?
  • Describe a local search algorithm for the image decomposition problem!
  • How would you implement this algorithm?
• Joint image decomposition and pixel classification
  • What do you know about the joint image decomposition and pixel classification problem?
  • How is pixel classification/image decomposition a special case?
  • Explain how semantic image segmentation is a joint image decomposition and pixel classification problem, and why neither pixel classification nor image decomposition alone are suitable abstractions!
  • Describe a local search algorithm for the joint image decomposition and pixel classification problem!
• Object recognition
  • What do you know about the object recognition problem?
  • How does a feasible solution to the object recognition problem look like?
  • Describe the objective function of the object recognition problem!
  • How are solutions affected in case an object is partially occluded?
  • How do solutions look like in case multiple instances of the same object are visible in the same image?
  • How are solutions affected by points that have a high cost for being any key point?
  • Can multiple such points be clustered together?
  • Can such a point be clustered together with points that are classified as key points?
  • How do you solve the object recognition problem?

Exercises

Assignments are published here as the course progresses.

For video conf discussions see post in OPAL.

• The first task is independent of the lecture. We use OpenCV for programming tasks to read, write and access images und provide us with a lot of Computer Vision code. We mostly use C++ in the exercise. It is necessary when you want to write Computer Vision libs, i.e. efficient code. Depending on your goals you may want to use Python.

  Slides and task are given in 1_IntroOpenCV_ss20.pdf

• In the next online exercise session on Tuesday, 14:50, I plan to discuss the last exercises. See Opal for code and comments. Here is the session link.

  From next week on I will use the lecture time slot (Friday, 3.DS, 11:10).

  I will also put a video about how to construct a graph for graph cuts (which was the topic of the last video conf) online - but only after Tuesday.