Martin Sundermeyer
@masundermeyer.bsky.social
1.1K followers 400 following 16 posts
3D Computer Vision & ML Research Scientist @Google
Posts Media Videos Starter Packs
masundermeyer.bsky.social
Seriously impressive demo by Boston Dynamics showing full body manipulation. This shows interaction with a complex environment rather than treating it as a disturbance like in all the robot dance videos.

youtu.be/HYwekersccY?...
Getting a Leg up with End-to-end Neural Networks | Boston Dynamics
YouTube video by Boston Dynamics
youtu.be
masundermeyer.bsky.social
Yesterday, our latest BOP Challenge report on model-based and model-free object pose estimation just received the Best Paper Award at the Computer Vision for Mixed Reality CVPR'25 Workshop :)

Check it out here: arxiv.org/abs/2504.02812
masundermeyer.bsky.social
Therefore, I'm quite excited to have docker-based robot grasping evaluations of SoTA pose detection algorithms from BOP!

The BOP challenge 2025 is still open for submissions. Present your results at our ICCV 2025 R6D workshop in Hawaii.
10th International Workshop on Recovering 6D Object Pose (R6D)
cmp.felk.cvut.cz
masundermeyer.bsky.social
6D pose detection metrics and grasp success rates strongly correlate, but they are not identical. Average Precision is a great threshold-free detection metric, but grasp success is often determined by the highest confident target object pose.
masundermeyer.bsky.social
At 4pm we will also give a talk on the early-bird results of the BOP challenge that now contains the same Industrial Plenoptic dataset and compare to the robotic BPC challenge (bpc.opencv.org).
BOP: Benchmark for 6D Object Pose Estimation
bop.felk.cvut.cz
masundermeyer.bsky.social
If you are at #CVPR_2025 today and into robotics, join our Workshop on Perception for Industrial Robotics Automation. We present the results of a joint pose estimation and grasping challenge on real robots, organized with #OpenCV and #Intrinsic, and announce 60k$ in prices.
CVPR 2025 Workshop - Perception for Industrial Robotics Automation (PIRA)
pira-workshop.github.io
masundermeyer.bsky.social
Does the recent progress in 3D vision transfer to challenging real-world problems in robotics and XR?

Proof it and participate in the BOP challenge 2025 featuring real-world datasets and tasks. bop.felk.cvut.cz/challenges/
masundermeyer.bsky.social
Got some recent research related to 6D Object Pose Estimation? Want to present it at #ICCV2025 in Hawaii? 🌴

Then submit and present at the 𝟏𝟎𝐭𝐡 𝐈𝐧𝐭𝐞𝐫𝐧𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐖𝐨𝐫𝐤𝐬𝐡𝐨𝐩 𝐨𝐧 𝐑𝐞𝐜𝐨𝐯𝐞𝐫𝐢𝐧𝐠 𝟔𝐃 𝐎𝐛𝐣𝐞𝐜𝐭 𝐏𝐨𝐬𝐞 (𝐑𝟔𝐃).

Paper deadlines: Jun 30 (in-proceedings), Aug 29 (non-proceedings)
10th International Workshop on Recovering 6D Object Pose (R6D)
cmp.felk.cvut.cz
masundermeyer.bsky.social
Real-time processing of sensor streams is crucial for robotics and AR. We introduce Troy-Vis, a real-time, open-vocabulary video instance segmentation method which will be presented in an oral presentation at WACV tomorrow.

Paper: arxiv.org/abs/2412.04434
Code: github.com/google-resea...
GitHub - google-research/troyvis
Contribute to google-research/troyvis development by creating an account on GitHub.
github.com
masundermeyer.bsky.social
In collaboration with Intrinsic and OpenCV we are running a new pose estimation + bin picking challenge that is decided by real robot grasp metrics. 🦾

Take the chance to bring your methods alive, win generous prices and present your approach at CVPR'25. 🏆
opencv.bsky.social
New Year, new Competition! The OpenCV Perception Challenge For Bin-Picking is a robotics and AI competition, focused on solving a real-world robotics problem and using real-life robot arms. Join a team and create the most accurate model to win a share of the $60k in prizes! youtu.be/kXsr5_v3Tho
OpenCV Perception Challenge for Bin-Picking Launch Video (Sponsored by Intrinsic)
YouTube video by OpenCV
youtu.be
masundermeyer.bsky.social
I'm still not used to generating realistic 4K videos in 1-2 minutes.. Glitches occur less frequently and physics are often imitated impressively well. #Veo2

youtu.be/w-lfkTrijv4?...
Veo2 Test // 100 Reasons
YouTube video by hellolaco
youtu.be
Reposted by Martin Sundermeyer
araffin.bsky.social
Publication-ready visualization of 3D objects and point clouds in seconds, using @blender.org and BlenderProc.

hummat.github.io/bproc-pubvis/
A screenshot of output from the blenderproc tool, with associated options. Left: Mesh Middle: Point Cloud Right: Depth
Reposted by Martin Sundermeyer
vaheta.bsky.social
1/ Excited to share that our latest work from Intrinsic will be presented as a paper at SIGGRAPH Asia 2024! 🎉

We’ve developed a plenoptic 3D vision system that addresses a key challenge in industrial robotics: providing robots reliable 3D input data. 🧵⬇️
masundermeyer.bsky.social
I'm open for new ways, like OpenReview on every arxiv paper with a community that creates incentives to collectively review all of them.
masundermeyer.bsky.social
Kaggle, Twitter, Discord and Github are great for mainstream and applied topics. But attention is not infinite and not fairly distributed. Peer review is not perfect but it ensures >0 closer looks on every idea, i.e. efficiently distributed attention.
masundermeyer.bsky.social
Interested to learn more?

Watch the recording of our ECCV 2024 workshop on Recovering 6D Object Pose.

Find the program and speakers here:
cmp.felk.cvut.cz/sixd/worksho...
9th Workshop on Recovering 6D Object Pose (R6D) - ECCV 2024
YouTube video by Meta Open Source
www.youtube.com
masundermeyer.bsky.social
Submitted a 6D Object Pose Estimation method at CVPR? 📝

Show the world that it actually works in practice and join the BOP challenge. 🦾

7 days left to win the BOP 2024 awards in the model-based and model-free tracks. 🏆
BOP: Benchmark for 6D Object Pose Estimation
bop.felk.cvut.cz