Polite Teacher
In construction site segmentation, like in many pixel-wise computer vision tasks, it is (relatively) easy to get the data. However, providing the labels is much more cumbersome. In this case not only because of the volume, but also due to specific construction expertise, which is required from the annotators.
As a consequence, one can have the “right” number of images for the training, but only a small fraction of them is labelled. Semi-supervised methods work exactly in such conditions.
We present Polite Teacher, a simple yet effective method for the task of semi-supervised instance segmentation. The proposed architecture relies on the mutual learning framework. To filter out noisy pseudo-labels, we use confidence thresholding for bounding boxes and mask scoring for masks. The approach has been tested with CenterMask, a single-stage anchor-free detector.
Tested on the COCO 2017 dataset, our architecture significantly outperformed the baseline at different supervision regimes. To the best of our knowledge, this was one of the first works tackling the problem of semi-supervised instance segmentation.
Bio
Dominik Filipiak works at Perelyn, a company devoted to solving complex AI problems in various industries. He’s a CEO and co-founder of a company branch devoted to computer vision, signal processing and robotics. He also teaches deep neural networks at the Faculty of Mathematics, Computer Science and Mechanics of the University of Warsaw.
In his scientific work, he explores modern methods of artificial intelligence, with a special focus on computer vision. He holds a MEng degree in Computing from Faculty of Computing at Poznań University of Technology and a PhD in Economics from Department of Information Systems at Poznań University of Economics and Business. Currently, he is doing his second PhD in Computer Science at University of Innsbruck.