This blog post is about our paper “Truly Multivariate Structured Additive Distributional Regression" published in the Journal of Computational and Graphical Statistics. You can access the full text here.
We extend the current component-wise boosting methodology for distributional copula regression to accommodate prevalent type of variables in biomedical research: bivariate binary, bivariate discrete, as well as mixed binary-continuous. One of the benefits of our boosting approach is its ability to handle high-dimensional input variables. We illustrate our proposed method by analysing three challenging datasets as well as a number of experiments.
Semi-supervised object detection (SSOD) reduces the need for extensive labeled datasets but faces real-world challenges such as class imbalance, label noise, and unreliable pseudo-labels. We investigate the impact of these challenges on SSOD and propose four novel data-centric building blocks—Rare Class Collage (RCC), Rare Class Focus (RCF), Ground Truth Label Correction (GLC), and Pseudo-Label Selection (PLS)—that improve SSOD frameworks. Our methods achieve up to a 6% performance increase by enhancing label quality and class representation while maintaining computational efficiency.
Together with the Continental Automotive GmbH, we developed an informed learning approach for deep learning motion prediction models in autonomous driving. Using probabilistic models to learn informative priors from synthetic training labels, we regularize the training on observed data, improving data-efficiency and robustness of state-of-the-art models.
WeedsGalore is a novel agricultural dataset, focusing on crop and weed segmentation from UAV imagery in croplands. The high plant cover (in terms of both number and area) and the availability of five spectral bands are key distinguishing characters of this dataset. When used with state-of-the-art deep neural networks, it leads to models that can effectively segment weeds, even in unseen fields.