Marc Van Droogenbroeck

Context . The SPHERE High-contrast Imaging survey for Exoplanets (SHINE) represents one of the largest and most sensitive direct imaging campaigns conducted to date with the VLT, targeting over 400 young, nearby stars with the goal of detecting and characterizing giant exoplanets and brown dwarfs. This extensive dataset offers a unique opportunity to revisit observations using modern, data-driven approaches, potentially uncovering new substellar candidates that may have been overlooked by classical analysis techniques. Aims . In this context, our study focuses on reprocessing and re-analysing the so-called F150 sample, a well-defined subset of 150 main-sequence stars within 100 pc observed in the H band with VLT/SPHERE as part of the SHINE survey. Methods . We applied NA-SODINN , a supervised deep learning model specifically tailored for detecting faint planetary signals in angular differential imaging sequences. Designed to model local noise properties by capturing spatio-temporal correlations, NA-SODINN produces network-confidence detection maps that push the detection of faint companions closer to the noise floor than classical S/N-based approaches. To translate the model’s pixel-wise confidence maps into actionable detections, we introduced a novel thresholding strategy that selects, for each annulus, the confidence threshold maximizing the F ₁ score (the harmonic mean of precision and recall) from injection-recovery experiments. This principled approach balances sensitivity and specificity, addressing a key limitation in current deep learning-based methods. Results . NA-SODINN recovers all known companions and some of the debris disks in the F150 sample, and it identifies 13 new substellar candidates not reported in previous studies: ten detected in both the H 2 and H 3 bands, and three in only one band. For the ten sources detected in both bands, we used the H 2– H 3 colour–magnitude diagram to perform a first assessment of their nature. Based on this analysis, we identified two ambiguous cases and three photometrically promising candidates. However, in light of the currently available multi-epoch SPHERE data, only the candidate around Smethells 20 remains a strong target for follow-up.

Text-to-image generative models can be used to imitate historical artistic styles, but their effectiveness in doing so remains unclear. In this work, we propose an evaluation framework that leverages expert knowledge from art history and visual semiotics and combines it with quantitative analysis to assess stylistic fidelity. Three experts rated both historical artwork production and images generated with Midjourney v6 for five major movements (Abstract Art, Cubism, Expressionism, Impressionism, Surrealism) and ten associated painters (male and female pairs), using nine visual criteria grounded in Greimas’s plastic categories and Wölfflin’s stylistic oppositions. Ratings were expressed as 95% intervals on continuous 0–100 scales and compared using our Relative Ratings Map (RRMap), which summarizes relative shifts, relative dispersion, and distributional overlap (via the Bhattacharyya coefficient). They were also discretized in four quality ratings (bad, stereotype, fair, excellent). The results show strong inter-expert variability and more moderate intra-expert effects tied to movements, criteria, criterion groups and modalities. Experts tend to agree that the model sometimes aligns with historical trends but also sometimes produces stereotyped versions of a movement or painter, or even completely missed its target, although no unanimous consensus emerges. We conclude that evaluating generative models requires both expert-driven interpretation and quantitative tools, and that stylistic fidelity is hard to quantify even with a rigorous framework.

Characterizing domains is essential for models analyzing dynamic environments, as it allows them to adapt to evolving conditions or to hand the task over to backup systems when facing conditions outside their operational domain. Existing solutions typically characterize a domain by solving a regression or classification problem, which limits their applicability as they only provide a limited summarized description of the domain. In this paper, we present a novel approach to domain characterization by characterizing domains as probability distributions. Particularly, we develop a method to predict the likelihood of different weather conditions from images captured by vehicle-mounted cameras by estimating distributions of physical parameters using normalizing flows. To validate our proposed approach, we conduct experiments within the context of autonomous vehicles, focusing on predicting the distribution of weather parameters to characterize the operational domain. This domain is characterized by physical parameters (absolute characterization) and arbitrarily predefined domains (relative characterization). Finally, we evaluate whether a system can safely operate in a target domain by comparing it to multiple source domains where safety has already been established. This approach holds significant potential, as accurate weather prediction and effective domain adaptation are crucial for autonomous systems to adjust to dynamic environmental conditions.