Visual object tracking
Visual object tracking is one of the main problems in computer vision. It has practical applications in: autonomous driving, human-computer interaction, film post production, image stabilization, measurements (e.g, sports), etc. I am going to give an overview of the state-of-art tracking algorithms, then focus on their practical implementation.
What is tracking – definitions, examples and applications. Two-view geometry. Two main approaches: tracking by motion estimation and tracking by detection. Lucas-Kanade tracker. Flock of trackers. Long-term tracking. Convolutional networks training and fine-tuning. CNN-based trackers. How to evaluate if your tracker is good. Accuracy, robustness and speed.
Project topics will be announced later.
The main programming tool for the course is (unfortunately) Matlab. The organizers will provide the students with a virtual machine with pre-installed software. But one can use Python for some projects.
The course is self-contained, but the student is expected to have a basic familiarity with calculus, statistics, machine learning and convolutional neural networks.
Work in computer vision since 2012. Commercial research: developed core technology for madora.co — personalized, visual based shopping app at Clear Research. Academy research: developed MODS – state-of-art algorithm for wide baseline stereo matching and LSUV – state-of-art method for convolution neural networks initialization.
Co-founder of Szkocka Research Group – Ukrainian open research community for computer science. Kaggler, 9th place out of 1049 teams at Kaggle National Data Science Bowl.
Open source contributor: deep learning libraries – caffe and keras. Reviewer for IEEE Transactions on Pattern Analysis and Machine Intelligence.
Field of interests: Local image features, deep learning, image matching, image understanding, remote sensing image interpretation.