Extends DINOv2, SAM and CLIP to 3D tasks by uplifting them in 3D with a simple learning-free rendering process. Collaboration with Thoth, Inria.

A UNIC classification model that distills from strong pretrained models, and performs on par of better then each of them. ECCV 2024

Automatically discover novel classes from unlabeled images and add the discovered classes to extend an existing detector. CoLLAs 2024

Task-specific distillation of large models such as DINOv2 or AVA. TMLR 2024

Weatherproofing Retrieval for Localization: universal retrieval model enhanced with synthetic images during training. ICLR 2024

Extension of EPIC KITCHENS dataset with 3D information. Collaboration with VGG, Oxford and MaVi, Bristol. NeurIPS 2023

Weakly Supervised Incremental Segmentation improved by leveraging the semantics underneath class names. CoLLAs 2023

Pretraining only using synthetic images obtained with Stable Diffusion. Collaboration with Thoth, Inria. CVPR 2023

Automatic data-augmentation for classification. Collaboration with Thoth, Inria. CVPR 2023

Supervised pretraining for strong results both on ImageNet and on transfer. Collaboration with Thoth, Inria. ICLR 2023

Unsupervised adaptation of SSL (eg DINO) to multiple retrieval tasks simultaneously. Collaboration with Naver Search. ECCV 2022

Fusing SSL (eg DINO) 2D representations with NeRF-like models using distillation. Collaboration with VGG, Oxford. 3DV 2022

A benchmark and a set of baselines for the realistic task of Online Adaptation for Semantic Image Segmentation. CVPR 2022

Unsupervised dimensionality reduction using a Barlow Twins loss. TMLR 2022

Mid-level features for image retrieval. ICLR 2022

Garnment image retrieval, querying with an example image and a textual modifier. ICLR 2022

A concept generalization benchmark for visual representations, named ImageNet-CoG. Collaboration with Thoth, Inria. ICCV 2021

Neural rendering as a way to segment out even seldomly moving objects in egocentric videos. Collaboration with VGG, Oxford. 3DV 2021

Learning Probabilistic Embedding for Cross-Modal Retrieval (PCME). Collaboration with Naver AI, CLOVA. CVPR 2021

Continual Adaptation of Visual Representations via Domain Randomization and Meta-learning. CVPR 2021 (oral).

Generating hard negatives in the feature space for improved self-supervised contrastive mearning. NeurIPS 2020.

Predicting masked tokens as a proxy task to learn visual representations from scratch using captioned images. ECCV 2020.

Fine-Grained Action Retrieval through Multiple Parts-of-Speech Embeddings. ICCV 2019.

Self-supervised Learning of Geometrically Stable Features Through Probabilistic Introspection. CVPR 2018 (spotlight).