First computer vision workshop

On January 14, researchers from the schools of the Institut Polytechnique de Paris organized the first edition of a workshop dedicated to computer vision, a key area of artificial intelligence. Gianni Franchi, teacher-researcher at ENSTA's Computer Science and Systems Engineering Unit, was a member of the organizing committee.

Bringing together the skills of 6 of France's top engineering schools, the Institut Polytechnique de Paris boasts a strong community of internationally recognized researchers in the field of computer vision, as demonstrated by the award of 70 million euros to the Hi! project last May. PARIS Cluster 2030 project.

Enzo Tartaglione, Gianni Franchi, Fanny Sabatier, Vicky Kalogeiton and David Picard organized the first workshop on the subject on January 14, 2025, with the aim of bringing this community to life and facilitating exchanges between researchers, teacher-researchers, doctoral students and industrialists.

Vicky Kalogeiton (École polytechnique) opened the sessions with a presentation of her work on multimodal image generation models, which can produce images from text, sound or visual data that has been interpreted by computer.

His presentation was followed by that of Enzo Tartaglione (Telecom Paris), during which he outlined his work on pruning techniques to make neural networks lighter and more energy-efficient, while minimizing performance loss.

Gül Varol (ENPC) continued with his work on computer estimates of 3D human motion following a text description of the desired modification.

David Filliat (ENSTA) followed with a presentation on computer vision for robotic tasks, before handing over to Renaud Marlet (ENPC) who presented his work on autonomous vehicles.

After a poster session and lunch break, the presentations resumed with Gianni Franchi (ENSTA) talking about robust algorithms in artificial intelligence in relation to uncertainty issues. His talk focused in particular on quantifying uncertainties in generative AI.

David Picard (ENPC) took over from him, presenting a new operation he has developed on how best to improve attention layers, a fundamental process in deep learning, so that they are more efficient.

He was followed by Titus Zaharia (Télécom SudParis) and his work with France Télévision on using AI to analyze images and generate subtitles and news banners.

 

Then Loïc Landrieu (ENPC) talked about attempts to produce a universal representation of geospatial data, before Christope Kervazo (Telecom Paris) closed the presentations with his interpretable and scalable deep learning methods for inverse imaging problems.

Most of these presentations came from publications in A*-ranked conferences, giving PhD students attending the workshop free access to them.

Given the success of the workshop, a second edition is planned for early 2026.