Researcher at @Google, 3D computer vision & machine learning. Previously PhD at ETH Zurich, intern at @Google, @Meta, @Microsoft, @magicleap.
Zurich, Switzerland
Joined January 2019
- Tweets444
- Following481
- Followers7.3K
- Likes2K
Excited to be at #ECCV2026 this week! Find me at
➡️ VidMap, a new system for 3D reconstruction from videos, poster #157 session #3 Thurs AM github.com/cvg/vidmap
➡️ ViLMa workshop Tuesday 8:45 for a talk covering VidMap and new works 🤫 vilma-workshop.github.io
See you there 🤩
Paul-Edouard Sarlin retweeted
Excited to share VidMap! #ECCV2026
We bring the temporal structure of SLAM into global SfM for accurate 3D reconstruction of long, unconstrained videos.
🚀 Code: github.com/cvg/vidmap
📄 Paper: arxiv.org/abs/2607.27194
🎥 Video: youtu.be/gy4szd3q5Oo
With @pesarlin & @mapo1
Paul-Edouard Sarlin retweeted
Pataki et al., "VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion"
SLAM uses temporal continuity but can drift; SfM can be out-of-order but ignores temporal continuity. In hindsight, we should obviously do both?
Paul-Edouard Sarlin retweeted
VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion
@zador_pataki, @pesarlin, @mapo1
tl;dr: RoMa v2 + MFT + DAv3 + MegaLoc + different handling of sequential tracking and loop closure + metric-scale GLOMAP with depth priors
arxiv.org/abs/2607.27194
Paul-Edouard Sarlin retweeted
🛰️ Introducing UniverSat: one transformer backbone for Earth Observation that handles ANY sensor, ANY spatial, spectral & temporal resolution, ANY scale — with a single set of weights. 🌍
Yes please! Open science FTW
📢📢📢 @CVPR motion proposal on best paper awards.
Motion: Eligibility for Best Paper Awards shall require the public release of both source code and model weights.
74%Support (require release)
26%Oppose (free market)
611 votes • Final resultsWhere is the code? No code, no public API, private datasets, too little details to reproduce = low impact. What does the award committee value?
Announcing the official #CVPR2026 Best Paper Award Winner! 🏆 Congratulations to the authors for their landmark contributions to the field!👏👏
Paul-Edouard Sarlin retweeted
IMW 2026 -- @pesarlin started talk with @taiyasaki tweet :)
Matas kicking of goated IMW26 @CVPR, come listen to keynotes by e.g. @jianyuan_wang and @pesarlin. Will have a 10 minute talk about LoMa-R around 4pm. See you in 504
Excited to be at #CVPR2026 this week 🤩 Catch me for 3 workshop talks:
➡️ E2E 3D learning, Wed PM
➡️ Image matching, Thur PM
Expect 🌶️ takes on learning/geometry and preview of new work.
➡️ EarthVision, Thur AM (+poster) about UniGeoCLIP, our latest work on geospatial learning.
Links:
e2e3d.github.io/
image-matching-workshop.gith…
grss-ieee.org/events/earthvi…
Let's grab coffee if you're attending too! I'm really looking forward to catching up with old friends, my last CVPR was already 3 years ago.
Paul-Edouard Sarlin retweeted
Feed-forward 3D reconstruction methods typically predict pointmaps in camera-centric frames. But why should a camera's arbitrary orientation define the coordinate system?
We introduce G3T, a transformer that predicts pointmaps in gravity-aligned frames. Regardless of input image orientation, our method always produces upright pointmaps (see demo).
We leverage this uprightness to create G3T-Long, a submap-based reconstruction method that improves robustness on long-sequence 3D reconstruction (more on that below).
Interactive demos, code, and model weights are available on our project page.
Paul-Edouard Sarlin retweeted
What if you could combine the precision of classical Structure-from-Motion with the local robustness of feedforward 3D models?
Meet GLUEMAP — a CVPR 2026 highlight ✨ SOTA across 5 datasets spanning 6 challenges.
w/ @ahojnnes & @mapo1
🌐website: lpanaf.github.io/cvpr26_glue…
Camera pose estimation is solved when one controls the hardware and can afford multiple cameras/IMUs. Examples: Google StreetView, Meta Aria, robotics (IMUs are so cheap). Monocular SLAM (no IMU) isn’t solved but has limited practical relevance, mostly internet/historical videos.
Paul-Edouard Sarlin retweeted
End-to-End 3D Learning @ CVPR 2026.
A workshop on moving from long 3D pipelines to trainable systems that map pixels or video directly to 3D.
Keynote talks include digital humans, scene generation, mapping, end-to-end VLA, and SLAM.
Paper deadline: May 1, 2026.
Best Paper Award: $500 cash + certificate
Best Demo Award: $500 cash + certificate
Website + CFP: e2e3d.github.io/
Paul-Edouard Sarlin retweeted
Excited to share my work as a Student Researcher at Google Zurich: UniGeoCLIP! 🌍🚀
W/ Eduard Trulls, Jan Hosang, @loiclandrieu & @pesarlin, we built a framework aligning 5 geospatial modalities in one space.
Presented at EarthVision @ #CVPR2026. 🧵👇
Paul-Edouard Sarlin retweeted
If only there were a benchmark for egocentric SLAM and VIO.... lamaria.ethz.ch/
We are publishing our first deep dive on what we believe is one of the most challenging layers in egocentric data - SLAM and VIO in the context of long-horizon state tracking.
We break down how SLAM and VIO fail in egocentric settings - visual features vanish at close range, depth sensors saturate, fast head motion blurs frames, and these failures don't always occur in isolation. They hit at the exact same moment, leading to compounding errors and making the downstream data unusable.
We believe the foundation for high-quality egocentric data demands sub-centimeter precision over long episodes ranging from a few minutes to up to an hour.
Paul-Edouard Sarlin retweeted
RaCo: Ranking and Covariance for Practical Learned Keypoints
Abhiram Shenoi @PhilippCSE @pesarlin @mapo1
tl;dr: ALIKED arch + DaD RL training with full 360deg rotaug with detector+covariance heads, separate ranker (sorter) model . No IMC eval
#3DV2026
openreview.net/forum?id=BWtd…
There are papers that I love and could spend hours reviewing (like this one, which I unfortunately didn’t get assigned to in peer-review), and others that I would rather not read. I like the bidding system of NeurIPS but it has never worked for me, I always get random papers!
github.com/zju3dv/EfficientL…
This is the depth of conversations between @pesarlin and Yifan Wang, one would dream to see in peer review. I'd dare to say, that is exactly peer review we want to have.
