Two papers accepted at CVPR 2024


Our papers "Task-Driven Wavelets using Constrained Empirical Risk Minimization" and "Kandinsky Conformal Prediction: Efficient Calibration of Image Segmentation Algorithms" have been accepted for publication at CVPR 2024!

"Task-Driven Wavelets using Constrained Empirical Risk Minimization" by Eric Marcus, Ray Sheombarsing and Jonas Teuwen was submitted to the subject area "Deep learning architectures and techniques". This paper constructs a theoretical framework for training a neural network with arbitrary constraints on its parameters. This is accomplished by computing the metric tensor of the Riemannian manifold that is a solution of these constraints, and subsequently performing stochastic gradient descent inside this manifold. Several example applications of the framework are presented, including an implementation where the filters of a convolutional neural network are forced to be wavelets, allowing for a multi-resolution approach to object delineation.

The second work, "Kandinsky Conformal Prediction", by Joren Brunekreef, Eric Marcus, Ray Sheombarsing and Jonas Teuwen addresses the issue of miscalibration in the context of image segmentation models. It is well-known that classifier models can output confidence scores that do not correspond to the accuracy of the model's predictions on unseen data. Such models can be calibrated using the framework of Inductive Conformal Prediction by making use of a held-back calibration dataset. Image segmentation models can be seen as a collection of many pixel classifiers, where each of these classifiers should in principle be calibrated separately. This makes proper calibration difficult, especially in low-data scenarios such as medical imaging. The Kandinsky method consists in grouping nearby pixels by their so-called "non-conformity curves", and simultaneously calibrating the resulting groups. This significantly increases the size of the available calibration sets. Experiments show that this indeed leads to improved calibration on a local level when the number of calibration images is small.