Seong Hun Lee

Postdoc working on 3D vision, SfM, Multiview geometry, etc.

[Zhenjun Zhao @zhenjun_zhao]

Non-Minimal Sampling and Consensus for Prohibitively Large Datasets

@SeongHunLee1, @patrick_vdwalle, @jcivera

tl;dr: seem like a non-minimal version of RANSAC

https://t.co/rAAl9YDORC

[Dmytro Mishkin 🇺🇦 @ducha_aiki]

Non-Minimal Sampling and Consensus for Prohibitively Large Datasets

Seong Hun Lee, Patrick Vandewalle, @jcivera
tl;dr: you split (huuuge) tentatives list, and run RANSAC of each. How do you combine winning hypothesis? Truncated Lp cost seems the best.
https://t.co/Zg6GlHGEnp

[Seong Hun Lee @SeongHunLee1]

@proximus is your internet down right now? I have no internet in Antwerp region! (Internet box)

[Zhenjun Zhao @zhenjun_zhao]

An Efficient Closed-Form Solution to Full Visual-Inertial State Initialization

@SamuelCerezoo, @SeongHunLee1, @jcivera

tl;dr: in title; not local solver; small-rotation and constant-velocity approximations->analytical solver->VI states

https://t.co/xfUK2nzz1R

[Seong Hun Lee @SeongHunLee1]

P3P Made Easy:
The Perspective-Three-Point (P3P) problem can be reduced to solving a polynomial. But the derivation is usually difficult to follow. We found a surprisingly easy solution that even high school students can understand (+SOTA performance!)
https://t.co/1rVO843avB

[Seong Hun Lee @SeongHunLee1]

In 7 hours, the Korean parliament is going to vote on the motion for impeaching President Yoon. If they fail to pass it (which may actually happen due to the mass delusion of the president's party), most likely we will soon witness a civil war. This insanity must stop now!!

[Seong Hun Lee @SeongHunLee1]

The Korean president is completely delusional and extremely dangerous. The declaration of emergency martial law today was a self-coup. It could've ended really badly. He is not just an idiot, but a delusional one who is willing to destroy and kill for himself and his wife.

[Seong Hun Lee @SeongHunLee1]

@ducha_aiki Thanks! I also enjoyed your presentation very much! I learned a lot from it! 😄

[Seong Hun Lee @SeongHunLee1]

Thanks for the photo! 😊

[Javier Civera @jcivera]

#ECCV2024 here we are! @SeongHunLee1 is presenting our works on point cloud registration and rotation averaging on the 6D object pose ws, amber 7+8

[Seong Hun Lee @SeongHunLee1]

Thanks for sharing our work!

[Dmytro Mishkin 🇺🇦 @ducha_aiki]

Alignment Scores: Robust Metrics for Multiview Pose Accuracy Evaluation

@SeongHunLee1 @jcivera

tl;dr: align (independently for translation and rotation) estimated and GT cameras. Do mAA-like histogram thing.
Pretty similar to what we used for IMC2024
https://t.co/cjdIoT9xbG

[Javier Civera @jcivera]

Feature Splatting shows better generalization than the original Gaussian Splatting for far-away novel view synthesis. We rank 2nd at the ScanNet++ challenge https://t.co/TSz8e0Ovv8 rendering at 82 fps, much faster than most methods in the benchmark including the 1st one

[I3A - UNIZAR @I3Aunizar]

#ConociendoA... @jcivera "El ritmo cada vez más acelerado de las innovaciones hace difícil prever los próximos desafíos"

👨🏻‍💻Pertenece al Grupo #RoPeRT y le encantaría ir de viaje a Japón, Australia y Argentina🗺

📰 https://t.co/D6vyscFH2s

[Min Choi @minchoi]

This is wild.

Simulon lets you add 3D models to your video in just few minutes.

1. 3D Robot with footsteps sound

[Michael Black @Michael_J_Black]

Young scientists regularly ask me for career advice. Academia or industry? Big company or startup? US or Europe?  Good scientists in AI disciplines are fortunate to have many choices. But choosing can be stressful. I always give the same advice. 1/10

[Vincent Sitzmann @vincesitzmann]

Introducing “FlowMap”, the first self-supervised, differentiable structure-from-motion method that is competitive with conventional SfM like Colmap!

https://t.co/ZwJGWfdytQ

IMO this solves a major missing piece for internet-scale training of 3D Deep Learning methods.

1/n

[Eric Brachmann @eric_brachmann]

📢A new learning-based approach to SfM: #ACEZero

No img-to-img matching, optimises image-to-scene correspondences directly. Needs no pose priors. Works on unordered image sets. Efficiently handles thousands of images.

Paper: https://t.co/Zc83YpoXpY
Page: https://t.co/kutgb81eFV

[Eric Brachmann @eric_brachmann]

✨ Presenting MicKey (#CVPR2024, Oral) ✨

We regress and match 3D camera coordinates rather then 2D key points, all in metric space. Gives you a scaled relative pose between two RGB images.

Paper: https://t.co/y3Ghwg6rM5
Page: https://t.co/DFsB60wH8D

[Seong Hun Lee @SeongHunLee1]

Wow, this is huge!
https://t.co/giffivQqqt

[Min Choi @minchoi]

Let's talk AI Music.

People are creating some amazing AI music and they are insanely good.

10 examples:

[Seong Hun Lee @SeongHunLee1]

Wow... Handling 10^7 point pairs with 99.8% outlier ratio? This is super impressive!

[Zhenjun Zhao @zhenjun_zhao]

Scalable 3D Registration via Truncated Entry-wise Absolute Residuals

Tianyu Huang, Liangzu Peng, René Vidal, Yun-Hui Liu

tl;dr: 3D registration->Truncated Entry-wise Absolute Residuals+3D&2D subproblems->tight lower and upper bounds->BnB

https://t.co/5ETIGIEF3Z