3D Structure From Visual Motion
This is a description page for the PhD course on 3D Structure from Visual Motion: Novel Techniques in Computer Vision and Autonomous Robots/Vehicles. This course can be taken also by students from Computer Engineering in the Laurea Magistrale track.
Course Aim & Organization
Simultaneous estimate of the unknown motion of a camera (or the vehicle this camera is upon) while reconstructing the 3D structure of the observed world is a challenging task that has been deeply studied in the recent literature. The PhD course on 3D Structure from Visual Motion: Novel Techniques in Computer Vision and Autonomous Robots/Vehicles will present modern techniques to simultaneously estimate the unknown motion of a camera while reconstructing the 3D structure of the observed world to be applied in scientific fields such as: 3D reconstruction, autonomous robot navigation, aerial/field surveying, unmanned vehicle maneuvering, etc.
Although formally entitled to just one of the teachers (myself) the course is also held by (in order of appearance)
With possibly special guests
Please consider this schedule as tentative ...
In the following you find the detailed schedule for the course and the rooms booked for it. In brackets you find also the lecturer for each specific topic.
- 3d Vision Basics (??? hours)
- Course introduction (1h M. Matteucci)
- Feature extraction, matching and tracking (2h M. Matteucci)
- Projection model and projection matrix (???h V. Caglioti)
- Fundamental and Essential matrices (???h V. Caglioti)
- Structure from Motion and Visual Odometry (??? hours)
- Optical flow (???h M. Marcon)
- Combined estiamation of 3D structure and camera egomotion (???h M.Marcon)
- Motion extraction and 3D reconstruction (???h V. Caglioti)
- Unconventional Visual Odometry (3 hours)
- Stereo e Omnidirectional odometry (TBD)
- Uncalibrated visual odometry (1.5h V. Caglioti)
- Omnidirectional odometry (1.5h V. Caglioti)
- Simulataneous Localization and Mapping (3 hours)
- From Bayesian Filtering to SLAM (1.5h M. Matteucci)
- EKF-Based SLAM (1.5h M. Matteucci)
- Visual SLAM (6 hours)
- EKF-based Monocular SLAM (3h D.G. Sorrenti)
- Why filters? PTAM and FrameSLAM (3h M. Matteucci)
- 3D without 3D (3 hours)
- Plenoptic methods, lumigraph, albedo, non Lambertian surfaces (3h M. Marcon)
Course Material & Referencies
The following is some suggested material to follow the course lectures.
Slides and lecture notes
- Camera geometry, single view, and two view geometry material
- Two view geometry and visual odometry material
- Correspondence analysis and RANSAC
- Optical flow tracking and egomotion estimation
- Structure from Motion
- Bayesian Filtering, Kalman Filtering, and SLAM
- Simultaneous Localization and Mapping a.k.a. SLAM!
- Monocular SLAM
- Stereo and Omnidirectional SLAM
- Panoramic Visual Odomentry
- Parallel Tracking and Mapping
- 3D without 3D
- R. Hartley, A. Zisserman. Multiple View Geometry in Computer Vision, Cambridge University Press, March 2004.
- S. Thrun, W. Burgard, D. Fox. Probabilistic Robotics, MIT Press, September 2005.
- Papers you might find useful to deepen your study:
- Simultaneous Localization and Mapping (SLAM): Part I The Essential Algorithms. H. Durrant-Whyte, T. Bailey 
- Unified Inverse Depth Parametrization for Monocular SLAM by J.M.M. Montiel, Javier Civera, and Andrew J. Davison 
- Parallel Tracking and Mapping for Small AR Workspaces by Georg Klein and David Murray 
- FrameSLAM: from Bundle Adjustment to Realtime Visual Mappping by Kurt Konolige and Motilal Agrawal 
Libraries and Demos
The course evaluation will be done on the basis of a project which could be completed also in groups of two people. In the case of PhD students this project could/should be somehow related to their research interests.