Meta learning could be defined as learning the parameters of the learning algorithm, which is useful in many realistic scenarios such as lack of training data, when requiring fast adaptation to the target domain/task, or in cases where one simply wishes to utilize the structure of the data more effectively. In this project, we study the viability of various meta learning algorithms and adapt them to real-world computer vision problems.


Learning to Adapt for Stereo.
Alessio Tonioni, Oscar Rahnama*, Thomas Joy*, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip H.S. Torr.
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.
[pdf] [supp] [arxiv] [code] [bib]

Similarity Learning for One-Shot Video Object Segmentation.
Mohammad Najafi*, Viveka Kulharia*, Thalaiyasingam Ajanthan, and Philip H.S. Torr.
CVPR Workshop: DAVIS Challenge on Video Object Segmentation, June 2018.
[pdf] [bib]