Get started
Work

L'adaptation de domaine dans l'ère des modèles de fondation

Year: 2025
Type: dissertation
Abstract: Artificial intelligence models that analyze images need to understand what each pixel represents, such as identifying cars and pedestrians in street scenes. Building such models for semantic segmentation typically requires collecting many manually labeled images, a time-consuming and expensive process. To address this bottleneck, domain adaptation has emerged: models are trained on accessible source data (such as synthetic images) and adapted to realworld target domains. However, this approach faces domain shift, where models trained on one data distribution fail to generalize to another at deployment. This challenge is especially pronounced in urban driving due to variations in weather, lighting, and geography. This thesis explores how Foundation Models, large AI systems pre-trained on millions of images, can overcome domain shift. These models learn powerful general-purpose representations that can be adapted to diverse applications. We investigate four practical scenarios under severe resource constraints. First, in one-shot domain adaptation, we generate synthetic datasets from a single target image by personalizing text-to-image diffusion models to capture target style while introducing diverse scenes. Second, we develop a collaborative framework where multiple foundation models work synergistically to generalize to unseen environments. Third, we distill knowledge from proprietary black-box AI services accessed via APIs without access to model internals. Finally, we improve open-vocabulary semantic segmentation by identifying which text descriptions work best for each object type. This work enables effective computer vision with unlabeled data and limited resources, democratizing access to advanced AI capabilities.
Language: English
Open Access status: green
OpenAlex ID: W7134153368
Cites:
Cited by:
Related to: