Tracking and Detection in Computer Vision - Winter Term 2019/20
Lecture by Dr. Slobodan Ilic
Exercises by: Mai Bui, Mahdi Hamad,Ivan Shugurov, Agnieszka Tomczak
Type: Lecture Module IN2210
Programs: Informatics (Bachelor, Master)
Biomedical Computing (Master)
Robotics, Cognition, Intelligence(Master)
ECTS: 7 Credits
Course Language: English
Time, Location & Requirements
Lectures: Mondays, from 21.10.19 at 14:00-16:00h at 00.13.009A.
Exercises: Thursdays, from 24.10.19 at 14:00-16:00h at 00.13.009A (max 70 participants), 03.09.012(20 places) and 03.11.018(20 places)
- The classes and exams are in English.
- For the midterm exam, nothing but your pens and your calculators are allowed.
- For the final exam, nothing but your pens, your calculators and one DIN A4 page (handwritten, front and back side) with notes are allowed.
- There is no repetition exam.
Computer Vision has emerged as a key discipline in computer science. Stemming from Artificial Intelligence(AI) with
the ultimate goal to understand Human Vision as one of the most powerful human senses it made a long way to what it is nowadays.
We are witnessing a big come back of Computer Vision to its roots "Artificial Intelligence" in the recent years. Recent boost in
Machine Learning and Deep Learning powered by incredible computational power of modern GPUs and profiting from massive amount of digital data clearly
changed the way how CV research is done at the moment and what impact it has on the society.
This is not only evident by a growing and highly competitive research community with a high impact
factor in computer science, but also by the emergence of numerous vision companies turning research
ideas into a myriad of commercial applications. Almost all big technology companies are at on the edge with Computer Vision, AI and machine learning
Besides well-known studies of 3D geometry and camera models, object tracking and detection in images and videos becomes one of the principal research
directions of modern Computer Vision. The main objective of this course is to provide students with a gradual introduction to object detection
and recognition techniques which proved to be widely used in the past and in current Computer Vision research.
The course will provide in-depth knowledge of image features, their detection and description,
matching techniques, key-point recognition, basic and advanced tracking algorithms based on image
features and image intensities. The course will address both hand crafted features and learned features with deep learning approaches.
So it will be mixture of established and emerging methods for object detection, recognition and pose estimation.
In the end, the students will have a thorough description of the most important object detection and pose estimation techniques. They should
be able to understand and implement those solutions and apply them in reasonably complex problems.
The concepts described in this course will be accompanied with brief explanations of the necessary
mathematical tools. The participants of this course will be given an important basis to follow the vast
and growing Computer Vision literature and use the acquired knowledge to solve new practical problems.
- First lecture will take place on 21.10. at 00.13.009A. Due to limited number of places it is recommended to come earlier to get a seat. After several lectures
usually there is enough available seats.
- First tutorial session will take place on 24.10 at 00.13.009A.
- Mid-term exam will take place on 05.12. from 18:30 bis 20:15h at Hörsaal MI-HS-1
The project will be done in MATLAB, C++ and Python and students are expected to do them on their own computer/laptop. They will consist of three projects that will include
one or several topics presented at the lectures. The projects will be presented to the students during the exercise sessions, where they will be instructed how to do the projects.
Each project is divided into a number of exercises and each of them corresponds to a homework which will be graded. In total all homework exercises will result in one exercise grade. The deadlines for the submissions of the results of the homeworks will be announced via Moodle.
Students have to make groups of up to FOUR students. If you are looking for a (TDCV) partner tell us and we will assign you one!
During the the presentation of the results for each project each group will have 5 min to present their solution in terms of slides, and also run the code live. This will be followed by questions from the tutors (~ 5min). All members of the group will be questioned about the algorithms and their implementation, so only those who can prove that they understood both theory and practice will get the respective points. Thus, having a working exercise doesn't mean that everyone in the group is going to get full points. Please note that only exercises which work might get the maximal number of points. There will be no "debugging-correcting"-sessions for your solutions during the project results presentation! If you have questions, ask them beforehand during the exercise sessions.
You can also find a short MATLAB tutorial here.
You are required to register via TUMOnline for the course. In case you are on the waiting list you will automatically be registered as soon as some of the registered students deregister.
We kindly ask registered students who do not intend to attend the course and take the exam to deregister and make space for those on the waiting list.
Who should attend this course?
Bachelor students in their last years and all Master students that are interested in learning about the modern and well-established concepts and algorithms
related to object detection, recognition and pose estimation in Computer Vision, or desire to use those techniques in their research.
Prerequisites: Most of the knowledge required should be part of the normal background in Computer
Science, undergraduate/graduate Mathematics and Geometry. However, strong knowledge in linear algebra and calculus will make much easier to follow the course.
Moreover, due to large part of the course is dedicated to project homeworks and implementations in MATLAB, C++ and python it is necessary to have good knowledge
of at least two of the programming languages.
For the geometry part of the course we recommend "Multiple-View Geometry" book of A. Ziserman and R. Hartely or respected papers attached to the slides, which will appear on this webpage.
For machine learning topics it is recommended to use the book of Christopher Bishop, Pattern Recognition and Machine Learning. And for deep learning, the Stanford course on CNNs available on-line.
. However, since the majority of the topics is state-of-the-art, students will be referred to the relevant publications which will be provided during the course and can be regarded as additional lecture notes.
Lecture Schedule and Content
Exercise Projects Schedule
If you have questions about a specific exercise, please direct them to the responsible tutor. At the due date the results of the projects will be presented by each group of students followed by the questions of the tutors. The students should be ready to run the code after their presentation as well as show parts of the code and be prepared to answer the questions about their implementation.