Detail of the student project

Topic:Deep learning for dense reconstruction from sparse depth measurements and RGB images
Supervisor:Doc. Ing. Karel Zimmermann Ph.D.
Description:Accurate 3D perception is an essential component for many fundamental capabilities such as emergency braking, predictive control for active damping, safe turning on a road intersection or self-localization from offline maps. Consequently, any fully-autonomous vehicle requires a sensor (e.g. Velodyne) providing high resolution and long range 3D measurements. The high resolution sensors are expensive, heavy, slow, and prone to mechanical wear, therefore low-resolution depth sensors (e.g. with only 4 row measurements planes) are often used in contemporary semi-autonomous cars. Autonomous estimation of high-resolution depth data from such sparse depth-measurements fused with RGB images seems to be a viable option for a close future. Learn a deep convolution neural network for dense depth reconstruction, given dataset (provided by thesis supervisor) captured by an autonomous car with calibrated (i) Velodyne sensor, (ii) cheap sparse depth-sensor and (iii) RGB camera.

Preffered qualification:
- B or better result achieved in a programming oriented subject or even better: the active participation in a programming competition (e.g. CTU Open Contest, ACM ICPC).
- B or better result achieved in a computer vision oriented subject.
- experience with Python and Tensorflow
- good mathemathical background.

supervisor's web:

Instruction:(1) Study state-of-the-art methods such as [1,2].
(2) Propose and implement you own algorithm.
(3) Evaluate proposed method on a selected dataset such as [3].
Bibliography:[1] Jiwon Kim, Jung Kwon Lee and Kyoung Mu Lee, 'Accurate Image Super-Resolution Using Very Deep Convolutional Networks', CVPR oral, 2016.
Max.number of students:6
Booked students:Dmitrii Noskov

Warning: the registration to the PTO can be canceled only by supervisor.
Responsible person: Petr Pošík