![cvprtum](/img/default-banner.jpg)
- 231
- 2 273 327
cvprtum
Приєднався 5 жов 2011
[ECCV 2024] Power Variable Projection for Initialization-Free Large-Scale Bundle Adjustment
European Conference on Computer Vision (ECCV), 2024
Authors: Simon Weber, Je Hyeong Hong, Daniel Cremers
Paper: arxiv.org/abs/2405.05079
Github: arxiv.org/abs/2405.05079
Abstract:
Most Bundle Adjustment (BA) solvers like the Levenberg-Marquardt algorithm require a good initialization. Instead, initialization-free BA remains a largely uncharted territory. The under-explored Variable Projection algorithm (VarPro) exhibits a wide convergence basin even without initialization. Coupled with object space error formulation, recent works have shown its ability to solve small-scale initialization-free bundle adjustment problem. To make such initialization-free BA approaches scalable, we introduce Power Variable Projection (PoVar), extending a recent inverse expansion method based on power series. Importantly, we link the power series expansion to Riemannian manifold optimization. This projective framework is crucial to solve large-scale bundle adjustment problems without initialization. Using the real-world BAL dataset, we experimentally demonstrate that our solver achieves state-of-the-art results in terms of speed and accuracy. To our knowledge, this work is the first to address the scalability of BA without initialization opening new venues for initialization-free structure-from-motion.
Authors: Simon Weber, Je Hyeong Hong, Daniel Cremers
Paper: arxiv.org/abs/2405.05079
Github: arxiv.org/abs/2405.05079
Abstract:
Most Bundle Adjustment (BA) solvers like the Levenberg-Marquardt algorithm require a good initialization. Instead, initialization-free BA remains a largely uncharted territory. The under-explored Variable Projection algorithm (VarPro) exhibits a wide convergence basin even without initialization. Coupled with object space error formulation, recent works have shown its ability to solve small-scale initialization-free bundle adjustment problem. To make such initialization-free BA approaches scalable, we introduce Power Variable Projection (PoVar), extending a recent inverse expansion method based on power series. Importantly, we link the power series expansion to Riemannian manifold optimization. This projective framework is crucial to solve large-scale bundle adjustment problems without initialization. Using the real-world BAL dataset, we experimentally demonstrate that our solver achieves state-of-the-art results in terms of speed and accuracy. To our knowledge, this work is the first to address the scalability of BA without initialization opening new venues for initialization-free structure-from-motion.
Переглядів: 64
Відео
[ACL 2024] Quality-Aware Language Models
Переглядів 103Місяць тому
Maximum-a-posteriori (MAP) decoding is the most widely used decoding strategy for neural machine translation (NMT) models. The underlying assumption is that model probability correlates well with human judgment, with better translations getting assigned a higher score by the model. However, research has shown that this assumption does not always hold, and generation quality can be improved by d...
[CVPR 2024] Cache Me if You Can: Accelerating Diffusion Models through Block Caching
Переглядів 6393 місяці тому
Project page: fwmb.github.io/blockcaching/ Arxiv: arxiv.org/abs/2312.03209 Authors: Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, Christian Rupprecht, Daniel Cremers, Peter Vajda, Jialiang Wang
[CVPR 2024] Sparse views, near light: a practical paradigm for uncalibrated Photometric Stereo
Переглядів 3623 місяці тому
Conference on Computer Vision and Pattern Recognition (CVPR), 2024 Publication: Sparse views, near light: a practical paradigm for uncalibrated point-light photometric stereo Authors: Mohammed Brahimi, Bjoern Haefner, Zhenzhang Ye, Bastian Goldluecke, Daniel Cremers. Paper: arxiv.org/abs/2404.00098 Abstract: Neural approaches have shown a significant progress on camera-based reconstruction. But...
[CVPR 2024] Flattening the Parent Bias: Hierarchical Semantic Segmentation in the Poincaré Ball
Переглядів 3433 місяці тому
Conference on Computer Vision and Pattern Recognition (CVPR), 2024 Publication: Flattening the Parent Bias: Hierarchical Semantic Segmentation in the Poincaré Ball Authors: Simon Weber, Baris Zöngür, Nikita Araslanov, Daniel Cremers Paper: arxiv.org/abs/2404.03778 GitHub: github.com/tum-vision/hierahyp Abstract: Hierarchy is a natural representation of semantic taxonomies, including the ones ro...
[CVPR 2024] Finsler-Laplace-Beltrami Operators with Application to Shape Analysis
Переглядів 3083 місяці тому
Conference on Computer Vision and Pattern Recognition (CVPR), 2024 Publication: Finsler-Laplace-Beltrami Operators with Application to Shape Analysis. Authors: Simon Weber*, Thomas Dagès*, Maolin Gao, Daniel Cremers (* equal contribution) Paper: arxiv.org/abs/2404.03999 GitHub: github.com/tum-vision/flbo Abstract: The Laplace-Beltrami operator (LBO) emerges from studying manifolds equipped with...
[ICML 2023] Robust Density-Aware Calibration
Переглядів 40911 місяців тому
[ICML 2023] Robust Density-Aware Calibration
[CVPR 2023] Results - Behind the Scenes: Density Fields for Single View Reconstruction
Переглядів 811Рік тому
[CVPR 2023] Results - Behind the Scenes: Density Fields for Single View Reconstruction
[CVPR 2023] Semidefinite Relaxations for Robust Multiview Triangulation
Переглядів 721Рік тому
[CVPR 2023] Semidefinite Relaxations for Robust Multiview Triangulation
[CVPR 2023] Behind the Scenes: Density Fields for Single View Reconstruction
Переглядів 2,5 тис.Рік тому
[CVPR 2023] Behind the Scenes: Density Fields for Single View Reconstruction
[CVPR 2023] Learning Correspondence Uncertainty via Differentiable Nonlinear Least Squares
Переглядів 838Рік тому
[CVPR 2023] Learning Correspondence Uncertainty via Differentiable Nonlinear Least Squares
[CVPR 2023] Power Bundle Adjustment for Large-Scale 3D Reconstruction
Переглядів 1,4 тис.Рік тому
[CVPR 2023] Power Bundle Adjustment for Large-Scale 3D Reconstruction
[ECCV 2022] Parameterized Temperature Scaling
Переглядів 760Рік тому
[ECCV 2022] Parameterized Temperature Scaling
DirectTracker: 3D Multi-Object Tracking using Image Alignment and Photometric Bundle Adjustment
Переглядів 1,6 тис.Рік тому
DirectTracker: 3D Multi-Object Tracking using Image Alignment and Photometric Bundle Adjustment
Intrinsic Neural Fields: Learning Functions on Manifolds
Переглядів 1,5 тис.Рік тому
Intrinsic Neural Fields: Learning Functions on Manifolds
The Probabilistic Normal Epipolar Constraint
Переглядів 1,4 тис.2 роки тому
The Probabilistic Normal Epipolar Constraint
[Code online] DM-VIO - ICRA2022 Presentation
Переглядів 2,6 тис.2 роки тому
[Code online] DM-VIO - ICRA2022 Presentation
[CVPR 2021] Post-hoc Uncertainty Calibration for Domain Drift Scenarios
Переглядів 6732 роки тому
[CVPR 2021] Post-hoc Uncertainty Calibration for Domain Drift Scenarios
DM-VIO: Delayed Marginalization Visual-Inertial Odometry [Code online]
Переглядів 11 тис.2 роки тому
DM-VIO: Delayed Marginalization Visual-Inertial Odometry [Code online]
TANDEM: Tracking and Dense Mapping in Real-time using Deep Multi-view Stereo
Переглядів 8 тис.2 роки тому
TANDEM: Tracking and Dense Mapping in Real-time using Deep Multi-view Stereo
ICCV 2021: Square Root Marginalization for Sliding-Window Bundle Adjustment
Переглядів 2,2 тис.2 роки тому
ICCV 2021: Square Root Marginalization for Sliding-Window Bundle Adjustment
ICRA'21 Presentation: Tight Integration of Feature-based Relocalization in Monocular Direct VO
Переглядів 1,6 тис.3 роки тому
ICRA'21 Presentation: Tight Integration of Feature-based Relocalization in Monocular Direct VO
CVPR 2021: Square Root Bundle Adjustment for Large-Scale Reconstruction
Переглядів 3,1 тис.3 роки тому
CVPR 2021: Square Root Bundle Adjustment for Large-Scale Reconstruction
MonoRec: Semi-Supervised Dense Reconstruction in Dynamic Environments from a Single Moving Camera
Переглядів 6 тис.3 роки тому
MonoRec: Semi-Supervised Dense Reconstruction in Dynamic Environments from a Single Moving Camera
[CVPR 2021] MonoRec: Semi-Supervised Dense Reconstruction in Dynamic Environments from One Camera
Переглядів 4,8 тис.3 роки тому
[CVPR 2021] MonoRec: Semi-Supervised Dense Reconstruction in Dynamic Environments from One Camera
TUM AI Lecture Series - Towards novel architectures for shape matching... (Maks Ovsjanikov)
Переглядів 1,7 тис.3 роки тому
TUM AI Lecture Series - Towards novel architectures for shape matching... (Maks Ovsjanikov)
[Presentation] LM-Reloc: Levenberg-Marquardt Based Direct Visual Relocalization
Переглядів 2,6 тис.3 роки тому
[Presentation] LM-Reloc: Levenberg-Marquardt Based Direct Visual Relocalization
[2 Min Summary] LM-Reloc: Levenberg-Marquardt Based Direct Visual Relocalization
Переглядів 7293 роки тому
[2 Min Summary] LM-Reloc: Levenberg-Marquardt Based Direct Visual Relocalization
Learning Monocular 3D Vehicle Detection without 3D Bounding Box Labels
Переглядів 2,2 тис.3 роки тому
Learning Monocular 3D Vehicle Detection without 3D Bounding Box Labels
4Seasons: A Cross-Season Dataset for Multi-Weather SLAM in Autonomous Driving
Переглядів 3 тис.3 роки тому
4Seasons: A Cross-Season Dataset for Multi-Weather SLAM in Autonomous Driving
thank yoy
The math here are succintly described. If I want to rigorously understand them, what textbooks should I look at? Could you please recommend?
Great work.
👍❤🌹Thank you so much for the extraordinary lectures and valuable knowledge of Prof. Cremers. A Chinese Learner. 2024-06-30.
Thank you, Prof. Cremers! I have finished the whole MVG lecture. It always makes me rethink the critical technique behind the method. And I have some new insight when I re-watch these classes. Thank you so much for the extraordinary lectures and valuable knowledge of Prof. Cremers. A Chinese Learner. 2024-06-30.
Thanks. Learned a lot even in 2024.
Thanks, It's very useful and very clear to revisit the basic algorithm and process of SLAM, and the optimization method behind it. Although it's 2024 now, I also found it a hard-core tutorial. A researcher marked this in 2024/06/21 Friday.
11:42 I wonder I got it right : by "preserving" , it doesn't mean cross product between 2 vectors is same before and after rigid body motion. Instead I think prof. meant applying a rigid body transformation(g) once after cross product of two vectors v and u is same as the cross product of vectors g(v) and g(u), where u and v is applied the same rigid body transformation g.
0:56 But I see a bunch of parallel edges relative to the camera, that were detected?
amrikai
He forgot to say: Please don't forget subscribing to the channel and liking the video 😂
ilk yorum
adamsin
türkler beğensin sayımızı bilelim
babageldi
Shoutout an alle Studenten die das gerade auch anschauen müssen
😇
Thank you so much for another wonderful lecture!
1:13:08
Hallo?
Und Live?
Und Live?!
Interesting. Thank you.
flat earth
Wow
This lecture is very clear and important and exhaustive. Danke.
Did it open sourced code?
Do you have slide?????
Great lectures!
Can we do this on a metaquest Pro headset?
I'm kind of confused about the difference between Fourier Features for Position encoding and autoencoders that use convolution + reconstruction loss. Does anyone understand why we prefer Fourier Features to convolutions in the NERF Case? Is it just because its a bag of pixels?
What a lecture! Thank you so much professer, this is so inspiring.
Can someone explain to me how the semi dense depth map is computed from two keyframes of a mono camera? I understand that pixels with image gradients are picked up, but how does one encode relative depth without say KLT tracking?
very nice work, thanks for your contribution to the SLAM society!
Is there a version for ROS2 ?
Thank you for these great lectures. Can the bag files of the exercise be made available ?
What is the significance of having a zero element in subspace ?
Great work! Can it work with equirectangular cameras such as Ricoh Theta V?
Very useful and explicitly, thanks for providing us with another way to understand the Lie algebra and Lie group in the image and camera scene. Thanks a lot. Hope us well in the pandemic.
Very useful and explicitly, thanks for providing us with another way to understand the Lie algebra and Lie group in the image and camera scene.
The transformations (marginalization operation in linear algebra) shown in (34:41) remove the 3D components in the equation and form the epipole constraint in 2D, but also lose some information since the process is not invertible. Thanks a lot for the demonstration!
Id like to try this. Are there any released software that I can try. I been using agismetashape and reality scan but it needs more data from different angles.
For more details on Lie theory I can recommend this video: ua-cam.com/video/nHOcoIyJj2o/v-deo.html
Awesome......
Awesome work! do you think this could work with a GoPro Hero 9 camera? comes with an IMU Bosch BMI260.
The IMU should be sufficient, but the camera probably has a rolling shutter which will decrease the accuracy. With a good camera calibration it might still be worth trying it out though.
@@cvprtum Thank you very much for your answer and for the code, can´t wait to try!
Nice, looking forward to trying on my devices
Danke Prof Cremers und Danke Mona. War auf jeden Fall sehr interessant!
Herr Cremers do be dripping tho
Cremers mag ich, Kombinatorik jedoch nicht..
Trotz technischer Probleme ein großes Danke an Prof. Cremers und das team im Hintergrund :)
WIR LIEBEN EUCH
sicke Vorlesung hat richtig spaß gemacht!! weiter so!
Welche Zentralübung/TÜ gehen zu diesem Vorlesung? ZÜ 8/TÜ9 hatten einige sachen, die in dieser 16ten VL nicht erwähnt wurden