Real-Time Optical Flow Using Neural Networks
From iis-projects
Contents
Short Description
Optical flow data is an ingredient to many complex computer vision systems. There are already quite a few algorithms to determine the optical flow between two frames very well. Some of these approaches use spatial correlation of small image patches, others work by looking at few well-trackable points (corners, edges, ...) in both images to find correspondences and then solve complex optimization problems to find reasonable solutions between these key points.
Evaluation of these algorithms is very time-consuming, often taking many seconds per frame on powerful workstations. This puts real-time applications out of reach where computer vision is most interesting: on low-power and mobile platforms, in cars, ...
We would like to take an unconventional approach to improve on this, training deep convolutional neural networks to calculate optical flow. The biggest issue of neural networks is usually to get a large enough, (hand-)labeled dataset. This is not a problem here, since we can take a high-quality algorithm's output as our reference input and getting raw video data without a lot of constraints is not really an issue.
As part of this project, the student(s) will:
- Explore existing optical flow algorithms
- Implement the most promissing candidate(s) (or make use of available implementations)
- Learn the fundamentals of convolutional neural networks (ConvNets)
- Get to know and use the Torch framework to train and adapt his ConvNet
- Evaluate the results (accuracy, performance)
- If there is time left: make a real-time demonstrator (camera interface, working platform, ... is already available)
The project can be adapted to the preferences of the student(s). Just drop me an email or stop by my office if you ..
- ... want to know more,
- ... are uncertain whether this project suits you, or
- ... if you have any other question.
Status: Available
- Looking for 1 Master or 2 semester project students
- Contact: Lukas Cavigelli
Prerequisites
- Knowledge of Matlab and/or C/C++
- Interest in video processing and machine learning
Character
- 30% Theory / Literature Research
- 70% software development
Professor
Detailed Task Description
Goals
Practical Details
Results
Links
[Motion tracking/optical flow using keypoint matching] [PhD thesis focusing on motion tracking/optical flow using keypoint matching (SIFT-Flow)] [Correlation-based, variational optical flow (don't worry about the math!)]