Iago Suárez
@IagoSuarez0
Followers
260
Following
402
Media
32
Statuses
107
ML Engineer at @Qualcomm XR Labs. Collaborating with CVG group of @ETH Zürich and PCR Lab of @La_UPM https://t.co/aAgPfUtl2Z
Madrid, Spain
Joined June 2021
📢I'm pleased to introduce GlueStick: A robust image matcher that combines points and lines📢. ➡️Website: ➡️Code: ➡️Paper: This is a collaboration with Rémi Pautrat, Yifan Yu, @mapo1 and @visionviktor. 1/🧵
5
21
89
Congrats to all my friends in @wayve_ai , this is a huge achievement 🔥🥳.
We’ve got some exciting news to share 🚨. We are thrilled to announce that we’ve raised $1.05bn in our latest fundraising round led by @SoftBank_Group with contributions from @nvidia and @Microsoft. This milestone will propel us in developing and launching our first Embodied AI
0
0
3
I just find this paper extremely cool! 😎. Actually, there's a very interesting overlap between this and GlueStick, because if both techniques were combined, it would be very easy to get low-poly meshes of the objects in the scene 🤔.
Code release for SuperPrimitives, and it comes with an interactive GUI! #CVPR2024 . SuperPrimitive is a new 3D representation which enables solving many 3D tasks at the level of image segments.
0
0
2
RT @RaquelUrtasun: On #InternationalWomensDay, I am feeling immense pride in the culture we've built at @Waabi_ai . Here's to all the incre….
0
6
0
RT @visionviktor: Interested in computer vision/ML/math/geometry?.Open PhD positions available in my group, more details at .
0
28
0
RT @VisionArtif: Mañana miércoles 31 de enero @IagoSuarez0 participa en los seminarios del @VisionArtif con la conferencia "Image matchin….
0
2
0
RT @ai_bites: Segment3D explores the use of 2D foundation models to automatically generate training labels for 3D segmentation. It predicts….
0
8
0
Congrats Dr. Peng @songyoupeng 🥳🎉 I am very proud of having a friend like you. You just did an incredible PhD!.
1
0
2
RT @milmadrid_: 📢 ¡Vuelve #MadridMLMeetup! Un encuentro para compartir los avanzes en #MachineLearning. 🗓️ 25/10 ⌚ 18:30 h.Con Iago Suáre….
0
4
0
💣We are releasing the training code for GlueStick in a joint effort with our friends from LightGlue! . With glue-factory you can train a matcher with any kind of feature, use different robust estimators, and evaluate in several benchmarks:.✅
github.com
Training library for local feature detection and matching - cvg/glue-factory
📢I'm pleased to introduce GlueStick: A robust image matcher that combines points and lines📢. ➡️Website: ➡️Code: ➡️Paper: This is a collaboration with Rémi Pautrat, Yifan Yu, @mapo1 and @visionviktor. 1/🧵
0
1
11
🚀Curious about the future of image matching? Today at 10:30 we are presenting GlueStick a joint deep matcher for points and lines in @ICCVConference! . Come by poster 021 in Foyer Sud and chat with Rémi Pautrar, me, Yifan Yu, @mapo1, and @visionviktor 😃
0
4
27
Very interesting talk about the future of robotics and SLAM by @AjdDavison at @ICCVConference 🤖
0
0
2
Takeaways🗒️:.➡️Efficient matching of points and lines simultaneously.➡️Unifying points and lines in a wireframe is essential to leverage the local connectivity of features.➡️ Geometry (Lines+Points) is an interesting alternative to dense matching.➡️SOTA results in multiple tasks.
1
0
0
We show SOTA results in several benchmarks for homography estimation, pose estimation (indoor & outdoor), and image stitching🚀. GlueStick doubles the quality of SOTA line matches (LineTR & SOLD2) and it gets better poses than LoFTR for InLoc although our matches are sparse!
1
0
0
We match wireframes using a Graph Neural Network (GNN) to combine each node's visual and spatial information. Here we use Self-Attention, Cross-Attention, and our novel Line Message Passing layers, which communicate nodes using the wireframe connectivity (details in the paper).
1
0
0
We extract points (SuperPoint), lines (LSD), and their descriptors. We combine them into a single wireframe joining line endpoints and keypoints that are nearby. Thus, we no longer match points & lines. Instead, we now match wireframes (graphs) from both images.
1
0
0
Contributions⚡️.- We propose a joint matcher for points and lines.- We replace previous heuristic geometric strategies for line matching with a data-driven approach.- We exploit the local connectivity between local features and unify points and lines into a wireframe
1
0
0