Detail of the student project

List
Topic:Učení odhadování optického toku bez supervize
Department: Skupina vizuálního rozpoznávání
Supervisor:Mgr. Jan Šochman, Ph.D.
Announce as:Diplomová práce, Bakalářská práce, Semestrální projekt
Description:We are part of the Toyota research lab, a research project at CTU sponsored by Toyota, and our goal is to understand dynamic scenes from a video recording, to know what moves and what is static, and where the moving objects move to. Our main application is a self-driving car equipped with a camera, but many other applications exist in robotics, human-computer interfaces, movie editing, ... Observing the scene through a single camera, however, one observes only a 2D projection of the actual 3D movement. We call this projection optical flow (OF) and it is represented as a 2D vector for every pixel in the frame.

The task of OF estimation has its roots in 80's and has been recently (as many other computer vision problems) dominated by solutions based on convolutional neural networks (CNNs). We have developed several algorithms in recent years ourselves, usually with world class performance. Yet, the problem is still not solved sufficiently well and even the best methods fail surprisingly often.

One of typical problems for CNN-based approachs is the lack of the data. In this project, we will take one of the state-of-the-art self-supervised methods [1], experiment with its training and the goal is to propose some interesting improvements based on the gained insight.

The exact problem formulation will be specified depending on the current needs of the project and the student experience (Bc./Ing.).

The student has to be able to code in Python and some knowledge of deep learning frameworks like PyTorch or Tensorflow is advantageous but not necessary.
Bibliography:[1] Unsupervised Moving Object Detection via Contextual Information Separation:
https://openaccess.thecvf.com/content_CVPR_2019/papers/Yang_Unsupervised_Moving_Object_Detection_via_Contextual_Information_Separation_CVPR_2019_paper.pdf
Responsible person: Petr Pošík