Yasser Benigmim profile photo

Yasser Benigmim

I am a PhD student in Computer Vision at Multimedia team (Telecom Paris) and VISTA team (LIX, École Polytechnique), supervised by Stéphane Lathuilière, Vicky Kalogeiton, and Slim Essid.

My PhD research focused on leveraging foundation models to advance generalization under distribution shift and limited supervision in semantic segmentation. I also interned at INRIA Paris in the Astra-vision team, where I worked on open-vocabulary semantic segmentation under the supervision of Raoul de Charette. I hold an engineering degree from École des Mines de Saint-Étienne and spent one year as an exchange student at EURECOM.

Email Scholar Github LinkedIn X

News

03/2026 Gave a talk at the Cohere Labs Open Science Community (YouTube).
02/2026 1st place in the Best Poster competition at MenaML Winter School 2026 (DGX Spark award).
02/2026 Released a reproducibility notebook for FLOSS: Colab, Jupyter notebook.
12/2025 I successfully defended my PhD thesis: Manuscript.
10/2025 Selected as top reviewer for NeurIPS 2025.
06/2025 Our paper FLOSS was accepted at ICCV 2025.
05/2025 Selected as outstanding reviewer for CVPR 2025.

Publications

ATGC teaser

Make me an Expert: Distilling from Generalist Black-Box Models into Specialized Models for Semantic Segmentation

Yasser Benigmim, Subhankar Roy, Khalid Oublal, Imad Eddine Marouf, Slim Essid, Vicky Kalogeiton, Stéphane Lathuilière
arXiv 2025

Introduces a black-box distillation setting in which segmentation models specialize from closed API outputs, with no access to weights or logits. DINOv2 attention entropy is used to identify informative scales for pseudo-labeling unlabeled images.

FLOSS teaser

FLOSS: Free Lunch in Open-vocabulary Semantic Segmentation

Yasser Benigmim, Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Raoul de Charette
ICCV 2025

Proposes a training-free method that uses prediction entropy to select per-class CLIP text templates, outperforming standard template averaging. As a plug-and-play module, it consistently boosts performance across open-vocabulary segmentation benchmarks.

CLOUDS teaser

Collaborating Foundation Models for Domain Generalized Semantic Segmentation

Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, Stéphane Lathuilière
CVPR 2024

Proposes a collaborative framework (CLIP + SAM + T2I Diffusion + LLM) for domain generalized semantic segmentation, without requiring any target data.

DATUM teaser

One-Shot Unsupervised Domain Adaptation With Personalized Diffusion Models

Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, Stéphane Lathuilière
CVPR-W 2023

Personalizes a diffusion model via DreamBooth on a single unlabeled target image to synthesize pseudo-target-domain data, enabling one-shot domain adaptation.

Misc

Awards

Academic services

Reviewing: I served as a reviewer for CVPR, ICCV, ECCV, NeurIPS and IJCV.

Teaching assistant:

  • Telecom Paris: Deep Learning (SD309), Linear Models (SD-TSIA204)
  • École Polytechnique: Computer Vision (INF649), Advanced Deep Learning (INF581A)

Talks

03/2026 Talk at Cohere Labs Open Science Community (YouTube).
06/2024 Talk at the GDR IASIS meeting (Slides).
03/2024 Talk at SLB AI Lab (Slides).