Lukas von Stumberg

PhD student at Computer Vision Group, TUM & Senior Computer Vision Engineer at Artisense

[Lukas von Stumberg @lukasvst]

We are interested in a wide range of areas including SLAM/VIO, SfM, Nerf, Gaussian Splatting, object detection, hand/eye/body tracking, and more. For this, Valve offers some of the best working conditions and benefits I have seen in our field.

[Lukas von Stumberg @lukasvst]

At Valve, we are looking for computer vision software engineers: https://t.co/LKaKITnqL5
If you want to create products for millions of customers and have proven experience in computer vision let me know!

[Lukas von Stumberg @lukasvst]

Tomorrow at 10:45 I will be presenting DM-VIO in the session Visual-Inertial SLAM (room 125) at #ICRA2022.

Below there's a sneak peak of the live demo I'll show during the poster session afterwards at pod 125, 5-8.

[Lukas von Stumberg @lukasvst]

I have just uploaded the ICRA2022 presentation video for DM-VIO: https://t.co/vJ2TgG0lP2
Check it out if you are interested in an explanation of Delayed Marginalization and Pose Graph Bundle Adjustment, which were proposed in our paper.

[Lukas von Stumberg @lukasvst]

The code for DM-VIO is now published: https://t.co/PIqqjCwK8F

Also, I am excited to share that I will present the paper at ICRA this year.

[Lukas von Stumberg @lukasvst]

Thanks to our novel IMU initializer based on delayed marginalization, our system works well on automotive datasets, while using only a monocular camera + IMU.

[Lukas von Stumberg @lukasvst]

I am excited to announce our new work "DM-VIO: Delayed Marginalization Visual-Inertial Odometry" which has been accepted at RA-L.
Paper: https://t.co/RORMDLKUBi
Code coming soon: https://t.co/PIqqjCwK8F
YT: https://t.co/KBRuhWtu2x
Page: https://t.co/ttFOBj8ozq

[Lukas von Stumberg @lukasvst]

We propose Delayed Marginalization. The idea is to maintain a second factor graph, where marginalization is delayed. This can be used for IMU initialization, and it also allows us to obtain an
updated marginalization prior by readvancing the delayed graph.

[Felix Wimbauer @felixwimbauer]

MonoRec is a novel method for dense reconstruction in dynamic environments from a single moving camera.

Proud that we will present at #CVPR2021!

CVPR Video: https://t.co/OfEgJc3swu
GitHub: https://t.co/QN4PlWzFjh
Project Page: https://t.co/kAQkuEKvle

#computervision #monorec

[Lukas von Stumberg @lukasvst]

@ddetone Thanks a lot, glad to hear that it makes a difference! :)

[Lukas von Stumberg @lukasvst]

We have just uploaded the ICRA presentation video for our work "GN-Net: The Gauss-Newton Loss for Multi-Weather Relocalization" to Youtube: https://t.co/cSOF43Q7bm

[Lukas von Stumberg @lukasvst]

@ducha_aiki @PatrickWenzelML That was the kind of comparison I was looking for. Very interesting paper, thanks for sharing!

[Lukas von Stumberg @lukasvst]

@ducha_aiki @PatrickWenzelML I have not checked in detail but I think PyRANSAC and DEGENSAC don't have an implementation for PnP, as there are no generate cases for this problem. In my experience the OpenCV implementation of solvePnPRansac works better than the Fundamental Matrix and Homography estimation.

[Lukas von Stumberg @lukasvst]

@ducha_aiki @PatrickWenzelML As far as I know the results are obtained with RANSAC + PnP, instead of RANSAC + Fundamental matrix estimation like in your results. It would be interesting to know if the differences you observe are also the case for PnP where the depths are known.

[Nan Yang @NanYang719]

Our new work --
"D3VO: Deep Depth, Deep Pose and Deep Uncertainty for Monocular Visual Odometry" is accepted as an oral presentation at CVPR 2020.
(1/n)
paper: https://t.co/CwCdT1ef6t

#CVPR2020 #cvpr #slam #computervison

[Lukas von Stumberg @lukasvst]

Thanks a lot! I am happy to announce that we will present this work next ICRA. I’m looking forward to meeting a lot of other scientists there!

[Tomasz Malisiewicz @quantombone]

GN-Net: The Gauss-Newton Loss for Multi-Weather Relocalization.

I presented this paper last week at my deep learning reading group. I love the idea of learning matchable representations via Gauss Newton image alignment. https://t.co/QgtRCc2x3X #ComputerVision #Robotics #slam