State-of-the-art Computer Vision
This course aims to give students a brief overview of state-of-the-art methods in computer vision (CV) so that when they are faced with a CV task they have a good understanding of how to approach it and what tools to use. This is done by reading, presenting and discussing impactful papers of the domain. On top of that we try our hand at different common CV tasks to give students practical experience.
Here is a non-final list of topics that will be covered in the course:
- Image classification
- Image segmentation
- Object detection
- Transformers for visual tasks
- Image dataset augmentations
- Methods for explainable AI
- Image denoising
- Contrastive learning
Prerequisites
We expect you to be comfortable with
- Linear algebra,
- (Matrix) calculus,
- Probability theory,
- Python (on a good level)
- Basics of machine learning
This is going to be a tough course, so do not take the prerequisites lightly.
Organization
We will use the reversed classroom approach, where we read papers at home and discuss them in the class. Discussions will be led by students themselves (we with instructors, of course, will be present). Depending on the final number of attendees, students will be divided into pairs or bigger groups, which will be responsible for moderating discussion for a particular topic.
The course is divided into four main blocks based on common CV tasks:
- Image segmentation
- Object detection
- Image classification
- Other tasks in CV: Image generation, denoising, augementations, model explainability, etc.
In each block we cover the most relevant papers to that task. Blocks are four weeks long and at the end of each block (excluding the last one) there is a Kaggle competition on the topic of the block, where students compete against baseline models and against one another. First three weeks of the block are for paper presentations and the last week is for discussing the results and approaches of the competition
There will be a lot of self-organization in this course, be ready!
Seminars are scheduled on Tuesday 16.15 - 18.00 in 1024 (most current info is on Slack), Delta (Narva mnt 18).
Zoom: link
We use Slack for communications.
Grading
To pass the course students need to
- Present twice in the seminars
- Take part in all three Kaggle competitions
Contacts
Joonas Ariva (joonas.ariva at ut.ee) and Pavel Chizhov (pavel.chizhov at ut.ee) are the main instructors of the course, Dmytro Fishman (dmytro.fishman at ut.ee) - is a responsible lecturer.