Uncertainty-Aware Transformers for Regression Tasks in Computer Vision

Convolutional Neural Networks (CNNs) and Transformers are currently the two dominating machine learning paradigms in computer vision due to their excellent performance in classification and regression tasks.

Both are variants of Artificial Neural Networks (ANNs), which, when trained for classification, use the activations in the final layer to encode both the class label and the corresponding confidence or uncertainty.

In contrast to the classification case, the uncertainty determination in regression networks, e.g. for pose estimation, is a largely unsolved problem.However, in many practical applications, both the regressed value and its confidence are required.

For instance in autonomous driving, it is not only relevant to know how far away a potential obstacle is located, but also how reliable this information is.

For CNNs, we proposed solutions in the previous VR project on normalized CNNs, but for transformers, this remains a challenge.

Therefore, this 4-years Ph.D. project aims to develop methods for determining the uncertainty in transformer-based regression networks for computer vision tasks.The research questions cover, besides the fusion of normalized CNNs and transformers, the choice of basis functions for the positional encoding (PE).

Suitable basis functions enable the probabilistic interpretation of relative PEs and provide means to estimate the uncertainty. The expected outcome is a more powerful approach to transformers for regression tasks in computer vision.