Bridging the domain gap for visual recognition systems.

Project description

Deep neural networks have significantly advanced modern vision recognition systems. However, performance of deep learning models depends on the assumption that the training and the test data are sampled from the same statistical distribution. In spite of their exceptional capabilities, deep learning models generalize poorly when exposed to new data distribution (or domains).

This issue aggravates further when test data contains classes unseen during training. While some recent works focus on merely segregating unseen classes as open set data, there are very few works that can simultaneously classify the in-domain classes and cluster the open set data into meaningful groups.

Furthermore, most works assume that the source data is present during the entire adaptation process. Such an assumption hinders the practical usage of vision systems which must adapt itself to new data continually, without access to the data on which it was previously trained. To circumnavigate the above challenges, we need to move towards open-set, potentially source-free, continual domain adaptation.

A very interesting use case of vision recognition is its application in Earth observation images. Earth observation images are acquired with a plethora of sensors that vary across spatial, spectral, and temporal resolution. Furthermore, Earth observation images vary geographically and over time. New classes appear and old classes keep disappearing depending on geographical aspects. Thus, we believe Earth observation images provide us a very interesting testbed for our open-set continual domain adaptation.

This project will have two major goals: 1) Methodologically improving open-set continual domain adaptation, in the context of adaptation both in presence and absence of source data; 2) Further optimizing the developed methods for Earth observation images and preparing suitable ready-to-use Earth observation datasets, if required. The project will contribute to progress the paradigm of Artificial Intelligence and Computational Geoscience.

Outcomes

Knowledge of machine learning, programming skill.

Essential capabilities

Knowledge of machine learning, programming skill.

Desireable capabilities

Knowledge of computer vision and deep learning toolboxes like PyTorch.

Expected qualifications (Course/Degrees etc.)

Masters in Computer Science or related field with course work in machine learning and/or computer vision.

Additional information for applicants

note: students must have own scholarship to apply (CSIR, UCG-NET, etc)