Kai Krajsek

Monocular depth estimation (MDE) is a challenging task in computer vision, often hindered by the cost and scarcity of high-quality labeled datasets. We tackle this challenge using auxiliary datasets from related vision tasks for an alternating training scheme with a shared decoder built on top of a pre-trained vision foundation model, while giving a higher weight to MDE. Through extensive experiments we demonstrate the benefits of incorporating various in-domain auxiliary datasets and tasks to improve MDE quality on average by ~11%. Our experimental analysis shows that auxiliary tasks have different impacts, confirming the importance of task selection, highlighting that quality gains are not achieved by merely adding data. Remarkably, our study reveals that using semantic segmentation datasets as Multi-Label Dense Classification (MLDC) often results in additional quality gains. Lastly, our method significantly improves the data efficiency for the considered MDE datasets, enhancing their quality while reducing their size by at least 80%. This paves the way for using auxiliary data from related tasks to improve MDE quality despite limited availability of high-quality labeled data. Code is available at https://jugit.fz-juelich.de/ias-8/mdeaux.

Augmentation-based self-supervised learning methods have shown remarkable success in self-supervised visual representation learning, excelling in learning invariant features but often neglecting equivariant ones. This limitation reduces the generalizability of foundation models, particularly for downstream tasks requiring equivariance. We propose integrating an image reconstruction task as an auxiliary component in augmentation-based self-supervised learning algorithms to facilitate equivariant feature learning without additional parameters. Our method implements a cross-attention mechanism to blend features learned from two augmented views, subsequently reconstructing one of them. This approach is adaptable to various datasets and augmented-pair based learning methods. We evaluate its effectiveness on learning equivariant features through multiple linear regression tasks and downstream applications on both artificial (3DIEBench) and natural (ImageNet) datasets. Results consistently demonstrate significant improvements over standard augmentation-based self-supervised learning methods and state-of-the-art approaches, particularly excelling in scenarios involving combined augmentations. Our method enhances the learning of both invariant and equivariant features, leading to more robust and generalizable visual representations for computer vision tasks.

To cope with the rapid growth in available data, the efficiency of data analysis and machine learning libraries has recently received increased attention. Although great advancements have been made in traditional array-based computations, most are limited by the resources available on a single computation node. Consequently, novel approaches must be made to exploit distributed resources, e.g. distributed memory architectures. To this end, we introduce HeAT, an array-based numerical programming framework for large-scale parallel processing with an easy-to-use NumPy-like API. HeAT utilizes PyTorch as a node-local eager execution engine and distributes the workload on arbitrarily large high-performance computing systems via MPI. It provides both low-level array computations, as well as assorted higher-level algorithms. With HeAT, it is possible for a NumPy user to take full advantage of their available resources, significantly lowering the barrier to distributed data analysis. When compared to similar frameworks, HeAT achieves speedups of up to two orders of magnitude.

The paper presents the first modelling experiment of the European-scale olive pollen dispersion, analyses the quality of the predictions, and outlines the research needs. A 6-model strong ensemble of Copernicus Atmospheric Monitoring Service (CAMS) was run throughout the olive season of 2014, computing the olive pollen distribution. The simulations have been compared with observations in eight countries, which are members of the European Aeroallergen Network (EAN). Analysis was performed for individual models, the ensemble mean and median, and for a dynamically optimised combination of the ensemble members obtained via fusion of the model predictions with observations. The models, generally reproducing the olive season of 2014, showed noticeable deviations from both observations and each other. In particular, the season was reported to start too early by 8 days, but for some models the error mounted to almost 2 weeks. For the end of the season, the disagreement between the models and the observations varied from a nearly perfect match up to 2 weeks too late. A series of sensitivity studies carried out to understand the origin of the disagreements revealed the crucial role of ambient temperature and consistency of its representation by the meteorological models and heat-sum-based phenological model. In particular, a simple correction to the heat-sum threshold eliminated the shift of the start of the season but its validity in other years remains to be checked. The short-term features of the concentration time series were reproduced better, suggesting that the precipitation events and cold/warm spells, as well as the large-scale transport, were represented rather well. Ensemble averaging led to more robust results. The best skill scores were obtained with data fusion, which used the previous days' observations to identify the optimal weighting coefficients of the individual model forecasts. Such combinations were tested for the forecasting period up to 4 days and shown to remain nearly optimal throughout the whole period.

This paper describes the pre-operational analysis and forecasting system developed during MACC (Monitoring Atmospheric Composition and Climate) and continued in the MACC-II (Monitoring Atmospheric Composition and Climate: Interim Implementation) European projects to provide air quality services for the European continent. This system is based on seven state-of-the art models developed and run in Europe (CHIMERE, EMEP, EURAD-IM, LOTOS-EUROS, MATCH, MOCAGE and SILAM). These models are used to calculate multi-model ensemble products. The paper gives an overall picture of its status at the end of MACC-II (summer 2014) and analyses the performance of the multi-model ensemble. The MACC-II system provides daily 96 h forecasts with hourly outputs of 10 chemical species/aerosols (O₃, NO₂, SO₂, CO, PM₁₀, PM_2.5, NO, NH₃, total NMVOCs (non-methane volatile organic compounds) and PAN+PAN precursors) over eight vertical levels from the surface to 5 km height. The hourly analysis at the surface is done a posteriori for the past day using a selection of representative air quality data from European monitoring stations. The performance of the system is assessed daily, weekly and every 3 months (seasonally) through statistical indicators calculated using the available representative air quality data from European monitoring stations. Results for a case study show the ability of the ensemble median to forecast regional ozone pollution events. The seasonal performances of the individual models and of the multi-model ensemble have been monitored since September 2009 for ozone, NO₂ and PM₁₀. The statistical indicators for ozone in summer 2014 show that the ensemble median gives on average the best performances compared to the seven models. There is very little degradation of the scores with the forecast day but there is a marked diurnal cycle, similarly to the individual models, that can be related partly to the prescribed diurnal variations of anthropogenic emissions in the models. During summer 2014, the diurnal ozone maximum is underestimated by the ensemble median by about 4 μg m⁻³ on average. Locally, during the studied ozone episodes, the maxima from the ensemble median are often lower than observations by 30–50 μg m⁻³. Overall, ozone scores are generally good with average values for the normalised indicators of 0.14 for the modified normalised mean bias and of 0.30 for the fractional gross error. Tests have also shown that the ensemble median is robust to reduction of ensemble size by one, that is, if predictions are unavailable from one model. Scores are also discussed for PM₁₀ for winter 2013–1014. There is an underestimation of most models leading the ensemble median to a mean bias of −4.5 μg m⁻³. The ensemble median fractional gross error is larger for PM₁₀ (~ 0.52) than for ozone and the correlation is lower (~ 0.35 for PM₁₀ and ~ 0.54 for ozone). This is related to a larger spread of the seven model scores for PM₁₀ than for ozone linked to different levels of complexity of aerosol representation in the individual models. In parallel, a scientific analysis of the results of the seven models and of the ensemble is also done over the Mediterranean area because of the specificity of its meteorology and emissions. The system is robust in terms of the production availability. Major efforts have been done in MACC-II towards the operationalisation of all its components. Foreseen developments and research for improving its performances are discussed in the conclusion.

In this paper, we examine the Ambrosio‐Tortorelli (AT) functional [1] for image segmentation from an estimation theoretical point of view. Instead of considering a single point estimate, i.e. the maximum‐a‐posteriori (MAP) estimate, we adopt a wider estimation theoretical view‐point, meaning we consider images to be random variables and investigate their distribution. We derive an effective block‐Gibbs‐sampler for this posterior probability density function (PDF) based on the theory of Gaussian Markov random fields (GMRF) [2].