Deep learning systems are demanding to train due to their inherent reliance on large amounts of annotated data. Acquiring annotations can be very costly, especially for fine-grained vision tasks such as semantic segmentation. Set aside time and monetary costs, the resulting datasets may be biased toward certain types of scene layouts and viewpoints, limiting recognition capabilities when the model operates under unfamiliar conditions, as also noted in earlier and contemporary work.
We have several activities on the topics of learning with little data, e.g.
- A post-doc project on "Embodied Visual Active Learning" financed by ELLIIT
- A PhD project (Olivier Moliner) on semi-supervised learning in collaboration with Sony
- A PhD project (Adam Tonderski) on contrastive learning in collaboration with Zenseact