Hermann Blum

📰Paper: arxiv.org/abs/2501.04597
🔥Code: github.com/cvg/Frontier...
🌍Page: boysun045.github.io/FrontierNet-...

w/ Boyang Sun, Hanzhi Chen, Stefan Leutenegger, Cesar Cadena, @marcpollefeys.bsky.social

We just released code, models, and data for FrontierNet!
Key idea 💡Instead of detecting froniers in a map, we directly predict them from images. Hence, FrontierNet can implicitly learn visual semantic priors to estimate information gain. That speeds up exploration compared to geometric heuristics.

Looking for a venue for your paper related to mapping / localization / retrieval / egocentric & embodied AI / domain shift / others ? Our ICCV 2025 workshop CroCoDL is still open for 8-page original submissions until the 30th of June via OpenReview: openreview.net/group?id=the....

Honestly ICCV2025 was the first time for me that I had great papers on my review pile that matched my interests & expertise. WAY better matching in my case than for the last 2 CVPRs.

ofc my standards are low coming from robotics where matching is based on keywords like „vision for robotics“ 🙈

This is the first time in a while I am creating a new talk. This will be fun!
I'll be up later today at the Visual SLAM workshop at @roboticsscisys.bsky.social‬

buff.ly/ADHxPsX

Special thanks to Tjark for creating this beautiful Lost & Found poster that we presented at the #CV4MR workshop.

Posters presented:
Guangda presented ARKitLabelMaker buff.ly/XcJHcz2

@haofeixu.bsky.social presented DepthSplat buff.ly/T0oWIdi

Dennis presented FunGraph, now accepted to IROS buff.ly/UvgZUzP

@zbauer.bsky.social‬ , @mihaidusmanu.bsky.social‬ and I presented CroCoDL buff.ly/ZHN2Ir4

Finally arriving home today after attending @cvprconference.bsky.social . This was the first #CVPR that I could attend in person! I expected it to be super crowded but was surprised - lots of time and space for chats at the poster session and the 15min talks could really go into detail.

Do you want to learn more about our novel dataset for Cross-device localization? Come by poster 121 and meet CroCoDL 🐊

cc @marcpollefeys.bsky.social @hermannblum.bsky.social @mihaidusmanu.bsky.social @cvprconference.bsky.social @ethz.ch

We just extended our submission deadline for 8-page paper submissions until June 30.
Accepted submissions go into ICCV WS proceedings 📄

all set for our poster lineup at the cv4mr.github.io #cvpr workshop

I‘ll be at CVPR this week and I am actively looking for PhD students (job announcement will go out the week after). Just send me a message if you are interested to meet up.

Excited to present our #CVPR2025 paper DepthSplat next week!
DepthSplat is a feed-forward model that achieves high-quality Gaussian reconstruction and view synthesis in just 0.6 seconds.
Looking forward to great conversations at the conference!

looking forward to it!

If you‘re watching #eurovision tonight, look out for the robots from ETH!
Really cool to see something I could work with during my PhD featured as a swiss highlight 🤖

We are organizing the 1st Workshop on Cross-Device Visual Localization at #ICCV #ICCV2025
Localizing multiple phones, headsets, and robots to a common reference frame is so far a real problem in mixed-reality applications. Our new challenge will track progress on this issue.
⏰ paper deadline: June 6

🏠 Introducing DepthSplat: a framework that connects Gaussian splatting with single- and multi-view depth estimation. This enables robust depth modeling and high-quality view synthesis with state-of-the-art results on ScanNet, RealEstate10K, and DL3DV.
🔗 haofeixu.github.io/depthsplat/

Exciting news for LabelMaker!
1️⃣ ARKitLabelMaker, the largest annotated 3D dataset, was accepted to CVPR 2025! This was an amazing effort of Guangda Ji 👏
🔗 labelmaker.org
📄 arxiv.org/abs/2410.13924

2️⃣ Mahta Moshkelgosha extended the pipeline to generate 3D scene graphs:
👩‍💻 github.com/cvg/LabelMak...

We are thinking about a bit similar setup and it seems you can record RGB-D really quite easily record3d.app What is the advantage of final cut pro?

*Please repost* @sjgreenwood.bsky.social and I just launched a new personalized feed (*please pin*) that we hope will become a "must use" for #academicsky. The feed shows posts about papers filtered by *your* follower network. It's become my default Bluesky experience bsky.app/profile/pape...

RA-L - IEEE Robotics and Automation Society Focus is on both applied and theoretical issues in robotics and automation. Robotics is here defined to include intelligent machines and systems; whereas automation includes the use of automated metho...

RA-L is a great model IMO: submit anytime, rapid-publishing (max 6 months including 1 month for revision), journal-style review process yields much better papers, accepted papers are automatically presented at the next ICRA/IROS.

www.ieee-ras.org/publications...

Open source code now available MASt3R-SLAM: the best dense visual SLAM system I've ever seen. Real-time and monocular, and easy to run with a live camera or on videos without needing to know the camera calibration. Brilliant work from Eric and Riku.

We have an excellent opportunity for a tenured, flagship AI professorship at @unibonn.bsky.social and lamarr-institute.org

Application Deadline is End of March.

www.uni-bonn.de/en/universit...

I don‘t doubt humanoids will get to where quadrupeds are right now, but my impression is that there is a gap of some years.

When I asked their sales team multiple times if they sell any humanoid robot that can walk stairs they would only confirm that the hardware is capable of that.

is it tough? When unitree showed their robots at RSS last year they could walk around a few steps on flat ground. I have yet to witness any of such agile movements in a real-world demo in uncontrolled environment, and then let‘s see if it is w/o safety harnish.

ORBSLAM3 is from 2021 and is still crazy good, but can run on even less than a MBPro.
We did some experiments lately using all the sota SfM stuff for loop closures and still on a number of recordings we could not beat ORBSLAM3

in all fairness, I have seen more CMT and openreview crashes at conference deadlines than papercept crashes
There are some advantages of a system that does not allow PDFs >10MB 😃

Very proud of Boyang for this work, great to see first shoutouts!

The gist is that exploration has always been treated as a geometric problem, but we show visual cues are really helpful to detect frontiers and predict their info gain.
W/ FrontierNet, you can get RGB-only exploration/object search/+

Just seeing this now, is there still space for me?