|Topic:||Perspectived object detection|
|Department:||Skupina vizuálního rozpoznávání|
|Announce as:||Diplomová práce, Bakalářská práce, Semestrální projekt|
|Description:||Box detection is a classic task of computer vision consisting in drawing a rectangular box around an object of interest (a car, a cat, a pedestrian). This is one way to represent the location and the nature of an object in the scene. This later allows high-level systems (e.g. autonomous car, ground robot) to understand the layout of the scene it is navigating in.
There is a significant amount of solutions that work very well (e.g. Faster R-CNN) and they are provided with open source code using easy-to-use libraries (e.g. Pytorch). One limitation of the current methods is that they detect only orthogonal boxes which do not fit accurately some objects. For example, a box can not fit a traffic sign seen with some amount of perspective. This is what motivates this project to design a box detector that outputs perspectives boxes.
The project consists in defining a parameterization of the box and training a neural network on a novel dataset that we recently collected (we provide access to GPUs) The resulting network will be integrated into our current research project related to visual localization. The student will also get to work in a research environment and interact with other students and researchers. Don't hesitate to write me if you have questions.
|Bibliography:|| Object-Guided Day-Night Visual Localization in Urban Scenes
 Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
 Oriented Object Detection in Aerial Images with Box Boundary-Aware Vectors