École Nationale Supérieure de
Techniques Avancées

Abstract

Artificial Intelligence is a field of computer science interested in creating systems capable of simulating behaviors typical to human intelligence: learning, generalization, understanding. Computer vision is a field of AI interested in using these techniques to interpret and understand the visual world, where models are trained with examples of images to be able to learn to detect objects and people, classify and segment images, generate new images, etc. A problem arrives when there are not many examples available of a certain anomalous category for us to train our model, detecting these anomalies is an important problem with many real world applications, like detecting animals on roads for autonomous vehicles, surveillance and security, we call this more generally the domain of out-of-distribution detection. The advances of computer power have allowed for more complex models to be developed, and more recently the field of NLP (Natural Language Processing) is intersecting more and more with computer vision, allowing the computer to learn from both text and image and thus have a more robust, general and deep understanding of scenes, objects, meaning and relationship. During this study we focused on applying vision-language models to the problem of detecting rare objects on image scenes.