4 Students

Join us

We are looking for strongly motivated students with an interest in applying machine learning to computer vision, robotics and more broadly artificial intelligence.

Internships and thesis projects can lead to a PhD in the ELLIS Unit at CIIRC in Prague with the possibility of spending part of the time at other Units in the ELLIS network.

Internship and Thesis topics

Analysis of Molecular Dynamic Simulations for Alzheimer’s Disease Research using VAMPnet Neural Networks

Supervisors
Jiri Sedlar, Josef Sivic, Tomas Pajdla, Torsten Sattler (CIIRC CVUT), Stanislav Mazurenko, Sergio Marques, David Bednar, Jiri Damborsky (LL MU)

Motivation
One of the characteristics of Alzheimer’s disease (AD) is the formation of neurotoxic aggregates of Aβ42 (Abeta) peptide. Understanding the dynamic properties of this protein is a key to determine the effects of drug candidates for potential AD treatment. Ambition of the project is to identify novel drug candidates for pre-clinical and clinical testing.

Read more >

Extreme Model Compression for Privacy Preserving Visual Localization

Supervisor
Torsten Sattler

Motivation
Visual localization is the problem of estimating the precise position and orientation, i.e., the camera pose, from which a given image was taken with respect to a known scene. Visual localization is a key component of many interesting applications, including self-driving cars and other autonomous robots and Augmented / Mixed / Virtual Reality systems. State-of-the-art visual localization approaches represent the scene using a Structure-from-Motion model, where each 3D point is represented by the local features from which it was triangulated.

Read more >

Motion Primitives for Reinforcement Learning in Robotics

Supervisors 
Vladimir Petrik, Josef Sivic

Motivation
Learning or fine-tuning robot motions to accomplish a given task successfully is becoming an alternative to the traditional robot programming paradigm. Reinforcement learning is usually used to train a policy represented by a neural network that predicts action (e.g., commands for the robot motors). Neural networks are simple to use, but their usage for robot motions is neither safe nor optimal. The danger of using simple networks lies in possible abrupt changes in actions that may result in dangerous high-frequency oscillations.

Read more >

Weakly-supervised learning of motion primitives from instructional videos

Supervisor
Josef Sivic

Motivation
This project is a step towards developing machines that can learn from humans how to perform different physical manipulation tasks. The learnt motion primitives should serve as an initialization for learning different robotic manipulation tasks transferable to different robotic platforms, e.g. the Humanoid robotic platform TALOS available at our partner lab in LAAS in Toulouse.

Read more >

Robotic Manipulator Planning, Control, and Learning with Object-Centric Feedback

Supervisor
Vladimir Petrik, Josef Sivic

Motivation
Every robot deployed in complex real world scenarios has to use motion planning and control capabilities in order to avoid obstacles and react to noise and disturbances. Our goal is to integrate the state-of-the-art image-based object pose estimation method and recently developed differentiable physics simulators into both planning as well as control. The integration with the planner will allow us to solve multi-step tasks with sequential dependencies while integration with control will increase the robustness and safety of the feedback-based control policy. Progress on these tasks will allow us to deploy robots into the noisy real world and achieve industrial-level performance with low-cost sensors.

Read more >

Efficient Camera Pose Hypothesis Verification Via Neural Rendering

Supervisors
Tomas Pajdla, Torsten Sattler

Motivation
Inspecting the quality of an estimated pose via a rendered image is an easy task for humans. However, existing quality metrics for machines struggle to achieve the same classification accuracy. As a result, wrong pose estimates lead to failure cases during 3D reconstruction or visual localization in applications such as autonomous robots and Augmented / Mixed / Virtual Reality, where failures can cause significant problems. This project is a step towards developing highly efficient verification techniques.

Read more >

Self-supervised learning for solving geometric construction problems from images

Supervisors 
Jiri Sedlar, Josef Sivic, Miroslav Olsak, Josef Urban

Motivation
This project is a step towards developing methods that combine machine learning with machine reasoning over noisy visual (or visual and text [Seo15]) inputs. The long-term goal is developing automatic “virtual AI assistants” that can help mathematicians with complex proofs [Hales17] or reason about text and associated illustrations (e.g. automatic patent lawyer assistant).

Read more >

Learning Local Features with Weak Supervision

Supervisor
Torsten Sattler

Motivation
Local features such as SIFT are a vital component of many 3D computer vision algorithms,including Structure-from-Motion, SLAM, and visual localization.  These algorithms rely on local features for data association, e.g., to establish correspondences between images or between images and 3D points. They fail if local features are not able to provide sufficiently many matches. In order to obtain features that are more robust to viewing condition changes, e.g., to changes in viewpoint or to changes illumination conditions such as day-night changes, deep learning …

Read more >

Learning to solve multiple-view geometry in Computer Vision

Supervisor
Tomas Pajdla

Motivation
We aim at using machine learning to address long-standing problems in multiple view geometry that traditional techniques cannot solve. For instance, current methods for computing camera geometry from image matches can cope efficiently with only relatively simple problems in two-view geometry, and there is still no efficient solver even for three-view geometry. We plan to develop a new approach to solving hard problems in multiple-view geometry by using machine learning to tune techniques from numerical algebraic geometry to the data, thus making them tractable and efficient.

Read more >