Detail of the student project

List
Topic:Double backpropation for feature detection network
Department:Katedra kybernetiky
Supervisor:Assia Benbihi
Announce as:Diplomová práce, Bakalářská práce, Semestrální projekt
Description:Feature extraction, description, and matching is a recurrent problem in vision
tasks such as Structure from Motion (SfM), visual Simultaneous Localization and
Mapping (SLAM) and scene recognition. The extraction consists of detecting image keypoints, then the matching step pairs the nearest keypoints based on their descriptor distance. Even though hand-crafted solutions, such as SIFT, have proven to be successful, recent breakthroughs on local feature detection and description rely on supervised deep learning methods. An alternative method, ELF [1] showed that a network trained on a standard task (e.g. classification) can be used as-is for efficient feature detection, without additional training. It exploits the correlation between the feature space and the image one: the gradient of the feature map with respect to the input image provides a saliency map with local maxima on relevant keypoint locations. The goal of this project is to train a neural network to optimize this correlation and improve feature detection.

[1] Benbihi, Assia, Matthieu Geist, and Cedric Pradalier. "ELF: Embedded
Localisation of Features in pre-trained CNN." In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pp. 7940-7949. 2019.
Bibliography:Benbihi, Assia, Matthieu Geist, and Cedric Pradalier. "ELF: Embedded
Localisation of Features in pre-trained CNN." In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pp. 7940-7949. 2019.
Responsible person: Petr Pošík