SupAereo is publicizing our prize AIAA paper on deep-learning based parameterization : isae-supaero.fr/.../best-scien… The paper itself can be found under infoscience.epfl.ch/entities/publi… .

In a paper to appear in TMLR, we show that we can train people trackers to estimate motion without motion annotations or re-identification algorithms: arxiv.org/abs/2411.16466 . The code is available github.com/cvlab-epfl/noi…. #deeplearning #computervision

If you are interested in the labeling of pulmonary tree structures, have a look at our recently published MIA paper sciencedirect.com/science/articl… . #deeplearn #medicalimaging #computervision

Check out my latest article: Non-Retirement and ERC linkedin.com/pulse/non-reti… via LinkedIn

Check out my latest article: Vision-Based Power Line Cables and Pylons Detection for Low Flying Aircraft linkedin.com/pulse/vision-b…

Check out my latest article: Temporally-Consistent Surface Reconstruction linkedin.com/pulse/temporal… via LinkedIn

Our paper, LeFusion: Controllable Pathology Synthesis via Lesion-Focused Diffusion Models, has been accepted by ICLR'25. Check out our code and models on GitHub: lnkd.in/es7hA_sg and the paper on arXiv: lnkd.in/ei8v876j.

I am back from visiting my “academic children” Xinchao Wang and Kaicheng Yu, now professors in Singapore and Hangzhou. They are all grown up and doing well ;-) And I am still flabbergasted at how much China has changed in the last three decades.

My Russian PhD student Andrey is defending his thesis today. He put up some bilingual slides. That's cute and I view this as one of the best answers we can give to unsavory politicians. Let's attract the best and brightest Russians to the west and then let's keep them around.

Supervised learning methods often assume that training and test data come from the same distribution. In a paper to appear in ICML, we show we can use idempotence to handle distribution shifts on-the-fly, without access to any additional labeled data. arxiv.org/pdf/2410.04201

Our paper written in collaboration with MBZUAI on modeling loose-fitting garments will be presented at SIGGRAPH. Crucially, we used guided diffusion models to synthesize the invisible parts while accurately modeling visible ones. arxiv.org/abs/2504.08353 #DeepLearning #Fashion

EPFL’s I&C department is doing well in the ERC EU research scene! actu.epfl.ch/news/ic-resear…

In paper to appear in MICCAI, we propose a method that “GenAI-fies” atlas-based segmentation by training a diffusion model to learn the joint distribution of images and segmentation masks. arxiv.org/abs/2503.06748 #deeplearning #Medical

A promising new approach to adapting models at inference time to be presented at ICML. linkedin.com/posts/durasovn… . The code is available online. #Deeplearning #computervision #NVIDIA #epfl

In a paper to appear at ICCV, we propose to use view-consistency distributions per-ray instead of fixed depth predictions to implement a sampling technique along the rays that implicitly regularizes the training of NeRFs. openreview.net/forum?id=ZQeKB…) #nerf

In a paper to appear at ICCV, we show how to count stacked objects such as those shown below, even though only those at the top are visible. Our key insight is that this suffices to estimate the number of objects per unit of volume. corentindumery.github.io/projects/stack…

In an upcoming MICCAI paper, we model the heart as interdependent SDFs representing a distinct anatomical component. By enforcing proper contact between adjacent ones, we ensure that they define anatomically correct structures. arxiv.org/abs/2307.08716

A View-consistent Sampling Method for Regularized Training of Neural Radiance Fields Aoxiang Fan, Corentin Dumery, Nicolas Talabot, Pascal Fua tl;dr: view-consistency distributions per-ray instead of fixed depth predictions->regularize NeRF training arxiv.org/abs/2507.04408

We released a large multi-view people-tracking dataset. We used 25 partially overlapping cameras placed over more than 400 meters. We acquired 24'000 images per camera, with full calibration data. scout.epfl.ch #VideoSurveillance #PeopleTracking #ComputerVision

The final version of our SIGGRAPH paper on 3D reconstruction of loose fitting garments from single images is now available on the ACM website dl.acm.org/doi/10.1145/37… . #deeplearning #computervision #GenerativeAI