On June 10th 2026, Estelle Chigot successfully defended a PhD thesis entitled Image-Based Synthetic-to-Real Domain Adaptation for Robust Object Recognition (abstract below). This industrial (CIFRE) thesis with Airbus was directed by Thomas Oberlin with my co-supervision and that of Meriem Ghrib and Manon Huguenin at Airbus. This thesis demonstrated how synthetic data could be used in computer vision pipelines in real applications, even the very challenging one of runway detection in adverse weather conditions. The theoretical and applicative contributions of this thesis demonstrated Estelle’s pragmatic and decidated approach to research, which was acknowledged by a Best of Track and 2nd Best of Conference Award at the conference DASC 2026. As of the writing of this post, Estelle is helping with teaching AI at ISAE-Supaero and is looking for research opportunities in computer vision and video games. I’m excited to see what she does next!
Image-Based Synthetic-to-Real Domain Adaptation for Robust Object Recognition
Deep learning models for object recognition require large annotated datasets, which are costly and difficult to obtain for industrial applications. Synthetic data generated by simulators offers a scalable alternative with perfect annotations, but models trained on synthetic images perform poorly on real data due to the domain gap. This thesis investigates image-based domain adaptation methods to bridge this synthetic-to-real gap, with a focus on semantic segmentation under adverse weather conditions. We first propose two diffusion-based style transfer methods, CACTI and CACTIF, that leverage semantic segmentation masks to transfer the visual style of real target images onto synthetic content while preserving structural coherence. Class-wise Adaptive Instance Normalization aligns feature statistics per semantic class, and selective attention filtering suppresses artifacts from weak cross-image correspondences. Experiments on the GTA5 to Cityscapes and ACDC benchmark show that these methods improve segmentation performance across difficult weather conditions. We then formulate the selection and ordering of style transfer operators as a multi-objective optimization problem, MOOSS, solved with evolutionary algorithms. Using paired-image metrics for efficient evaluation, the method explores the trade-off between content preservation and style similarity. While the optimized pipelines produce visually diverse results, they reveal that current perceptual metrics do not reliably predict downstream task performance. Finally, we apply these approaches to runway detection for automatic landing, an industrial use case where real data is scarce and domain shift is significant. Experiments confirm that style transfer improves robustness to adverse conditions, while also exposing the limitations of current benchmarks dominated by urban driving data.