École Nationale Supérieure de
Techniques Avancées

Résumé

The categorization of environments for domain adaptation in computer vision is essential for enhancing the functionality of mobile robots in diverse settings such as roads, pathways, and forests. This project aims to develop representation spaces that accurately characterize these environments, facilitating the automatic adjust- ment of computer vision algorithms for tasks like detection and depth prediction based on the current context. The methodology involves adapting metric techniques for anomaly detection, utilizing norm measurements in residual spaces. A notable approach includes em- ploying the space generated beyond the primary principal components, derived either from image patches sampled in various environments or from a latent space of deep features. This weakly supervised technique eliminates the need for training a classifier on different environments, thus enhancing efficiency. Initial experiments will leverage the CBIR 15-Scene dataset, which is specifi- cally designed for environment categorization. The outcomes of this research will contribute to the development of more adaptive and context-aware computer vi- sion systems for mobile robots. The code can be found here: ⟨https://github.com/ sergio-contente/PRE 2024⟩