iiinova — where your digital self keeps evolving

Eric Brachmann @ericbrachmann.bsky.social · 24/09/2026

What fresh hell is this? Scholar inbox now comes with 2 min podcasts of each paper... I randomly choose two relocalization papers and they both started with "Imagine navigating a large, dark warehouse" ... No - I do not want to imagine this. Was this trained on YouTube slasher flicks? Why "dark"?!

181

Eric Brachmann @ericbrachmann.bsky.social · 03/09/2026

The headline numbers of AutoCompass are just amazing and set a new SOTA for this task. By switching to weak supervision, you beat associated baselines. Also some interesting findings: train on coarse position without heading labels only, and accurate heading prediction just comes as a side product.

100

Eric Brachmann @ericbrachmann.bsky.social · 03/09/2026

So we train with weak supervision instead. The result is a set of losses catering to different training settings: Got images with GPS but no heading? Check. Relative poses but not scale-metric? Sure. Scale-metric relative poses? Great, bring them. Even coarsely geo-aligned? We have a loss for that.

100

Eric Brachmann @ericbrachmann.bsky.social · 03/09/2026

To train these systems, you need many examples of images with geo-located pose annotations. But the ground truth of existing datasets is unreliable and often imprecise. At the scale required for training, better ground truth is difficult to obtain.

100

Eric Brachmann @ericbrachmann.bsky.social · 03/09/2026

Need an accurate camera pose, but all you have is an image and a coarse location, e.g. from GPS? No problem, just localize against OpenStreetMap. Methods unlocking that capability are called Neural Map Matchers which had poor accuracy so far. But we found the problem.

110

Eric Brachmann @ericbrachmann.bsky.social · 03/09/2026

New #ECCV2026 paper alert: 🧭 AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels 🧭 Paper: arxiv.org/abs/2609.02798 Website: nianticspatial.github.io/autocompass/ Video: www.youtube.com/watch?v=hmUF...

1134

Eric Brachmann @ericbrachmann.bsky.social · 09/06/2026

ACTUALLY, that very same asteroid was already discovered 1990 by the hobby astronomer, Jürgen S.

3775

Eric Brachmann @ericbrachmann.bsky.social · 20/05/2026

I have had 70 co-authors so far. I took some time to wire all these names up with their websites on my personal page. (The BOP papers have been a real pain. 😅) Fun to browse through what all these people have been up to since we collaborated.

170

Eric Brachmann @ericbrachmann.bsky.social · 25/03/2026

Thanks, but no thanks. You might get me after all if we submit, but for now it feels good to decline. Last year, it was a horrible experience to review for NeurIPS. And I had a reduced load of two papers....

1111

Eric Brachmann @ericbrachmann.bsky.social · 06/03/2026

"Authors should not use negative v-spaces to change the template layout." The template layout:

0285

Eric Brachmann @ericbrachmann.bsky.social · 04/02/2026

It's time, dear #CVPR2026 reviewers. Time to decide.

1153

Eric Brachmann @ericbrachmann.bsky.social · 28/01/2026

Interviews for our 2026 research internships will commence soon. It is not too late to apply! A lot of exciting stuff is happening in 3D vision, let's explore it together.

0152

Eric Brachmann @ericbrachmann.bsky.social · 17/12/2025

What a nice morning to find that Alzugaray, Taher @marwantaher.bsky.social , Davison @ajdavison.bsky.social have turned ACE into a SLAM system. Since we did ACE, I was wondering whether it was possible, and I always thought who if not those folks can pull it off. 🙂 ACESLAM: arxiv.org/abs/2512.14032

1313

Eric Brachmann @ericbrachmann.bsky.social · 10/11/2025

‼️ Major update to the #ACEZero code base. We added the capabilities of the #ICCV2025 paper on Scene Coordinate Reconstruction Priors. Gives you, e.g., a RGB-D version of ACE0. In the snipped below, I jointly reconstruct multiple RGB-D scans of ScanNet. github.com/nianticlabs/...

1112

Eric Brachmann @ericbrachmann.bsky.social · 17/10/2025

Looking forward to a busy #ICCV2025. I will give three (very different) talks at workshops and tutorials, see info below. We also present two papers, ACE-G and SCR Priors. And it's the 10th (!) anniversary of the R6D workshop, which we co-organize.

1124

Eric Brachmann @ericbrachmann.bsky.social · 15/10/2025

🌟 Scene Coordinate Reconstruction Priors at #ICCV2025 🌟 Can we learn what a successful reconstruction looks like, and use this knowledge when reconstructing new scenes? Explainer: youtu.be/RkV6U5xYc20 Project Page: nianticspatial.github.io/scr-priors Inspiring work by Wenjing Bian et al.!

1202

Eric Brachmann @ericbrachmann.bsky.social · 14/10/2025

🔥 ACE-G, the next evolutionary step of ACE at #ICCV2025 🔥 We disentangle coordinate regression and latent map representation which lets us pre-train the regressor to generalize from mapping data to difficult query images. Page: nianticspatial.github.io/ace-g/ Stellar work by Leonard Bruns et al.!

2245

Eric Brachmann @ericbrachmann.bsky.social · 02/10/2025

FastForward was trained on challenging frame sets. Therefore, it is able to solve difficult relocalization cases, like opposing shots in the video below. We only have mapping images of the front of the sign. Still FF relocalizes well even behind the sign.

020

Eric Brachmann @ericbrachmann.bsky.social · 02/10/2025

FastForward uses 20 retrieved images by default. But no need to compute attention over 21 images. To keep the model lean, we sub-sample the feature tokens of mapping images. Each token is combined with its ray embedding. Hence, FastForward operates on samples of visual tokens in scene space.

110

Eric Brachmann @ericbrachmann.bsky.social · 02/10/2025

Building visual maps via reconstruction (SfM, SCR) can take time. Instead: ⏩ Get poses for your mapping images, e.g. via SLAM in real time. ⏩ Build a retrieval index in seconds. ⏩ FastForward predicts the relative pose between a query image and all retrieved images in one forward pass.

200

Eric Brachmann @ericbrachmann.bsky.social · 02/10/2025

⏩ FastForward ⏩ A new model for efficient visual relocalization. Accurate camera poses without building structured 3D maps. nianticspatial.github.io/fastforward/ Work by @axelbarroso.bsky.social, Tommaso Cavallari and Victor Adrian Prisacariu.

1283

Eric Brachmann @ericbrachmann.bsky.social · 29/09/2025

Small but important quality of life update by @rerun.io: Customizable frustum colors. Finally we can distinguish estimates and ground truth in the same 3D view ;)

1172

Eric Brachmann @ericbrachmann.bsky.social · 26/09/2025

Recovering 6D Object Pose - Looking forward to the R6D workshop at #ICCV2025 in Hawaii! (not looking forward to the jet lag) cmp.felk.cvut.cz/sixd/worksho... The accompanying challenge is still open for submissions until Oct 1st. And the workshop also still accepts non-proceeding paper submission!

041

Eric Brachmann @ericbrachmann.bsky.social · 03/09/2025

Did you know in Germany, there are signs in every city indicating the way to the closest H100, in case of an upcoming deadline?

3302

Eric Brachmann @ericbrachmann.bsky.social · 25/07/2025

Sorry folks, #ICCV2025 booked out for vegetarians.

030

Eric Brachmann @ericbrachmann.bsky.social · 01/07/2025

How about a fun game of career path domino instead?

081

Eric Brachmann @ericbrachmann.bsky.social · 23/06/2025

Thank you, but I think your talk was awesome :) And here for all to enjoy again: eccv.ecva.net/virtual/2024...

010

Eric Brachmann @ericbrachmann.bsky.social · 23/06/2025

I was not aware that the #ECCV2024 oral recordings are publicly available... So here is the #ACEZero talk: eccv.ecva.net/virtual/2024...

3282

Eric Brachmann @ericbrachmann.bsky.social · 27/05/2025

This is from the #ICCV2025 workshop on Camera Calibration and Pose Estimation (CALIPOSE), organized by @sattlertorsten.bsky.social et al. Now I have 5 months to get used to the fact of giving a talk after Richard Hartley 🙈

1210

Eric Brachmann @ericbrachmann.bsky.social · 08/04/2025

The BOP challenge 2025 is already running: bop.felk.cvut.cz/challenges/b... Your chance to shine on the task mentioned above. But beware that BOP is one step ahead already, by adding another set of nightmarish datasets. 👻

041

Eric Brachmann @ericbrachmann.bsky.social · 08/04/2025

The BOP report of 2024 is out: arxiv.org/abs/2504.02812 BOP is a benchmark for 6D object pose estimation, with a new challenge organised yearly. Community: Model-based 6D localisation? Pff, easy-peasy. 💪 BOP: Ok. Model-free 6D detection on the datasets below 👇 Community: 🙈 (0 entries submitted)

1194

Eric Brachmann @ericbrachmann.bsky.social · 04/04/2025

I do not think so. You get the same moment vector for different choices of the reference point along the line. See e.g. faculty.sites.iastate.edu/jia/files/in...

120

Eric Brachmann @ericbrachmann.bsky.social · 10/03/2025

Back to healthier habits, after the deadline. </promotion>

270

Eric Brachmann @ericbrachmann.bsky.social · 10/01/2025

I sometimes imagine these reviewers as people standing before a piece of abstract art and claiming "I could have done that!" (But you didn't.)

Person standing before a painting in a museum: "I can do that!". The painting says: "No, you can't."

0110

Eric Brachmann @ericbrachmann.bsky.social · 28/11/2024

Recently, we trained the MicKey matcher with DSAC, and that worked more or less from scratch. "More or less" because MicKey uses a DINOv2 backbone, which also is a good init I guess. nianticlabs.github.io/mickey/

050

Eric Brachmann @ericbrachmann.bsky.social · 28/11/2024

Posting some evergreens for the new crowd. Did you now you can differentiate RANSAC? If you fix the # of iterations, RANSAC is an argmax over hypotheses. You turn the inlier count into your policy for hypothesis selection, and train with policy gradient (DSAC, CVPR17). github.com/vislearn/DSA...

2787

Eric Brachmann @ericbrachmann.bsky.social · 26/11/2024

Updated my website with all the current social media options... I guess I will keep checking Twitter lazily, and post big announcements (similar to LinkedIn), and rely on Bluesky for discussions and chit-chat. At the moment, Threads does not seem to be worth putting any effort into.

0110

Eric Brachmann @ericbrachmann.bsky.social · 19/11/2024

It's been a while since I gave a proper lecture, but this Thursday I'll have the honour with a guest lecture in the geometry class of CMU. From the beginnings of scene coordinate regression all the way to #ACEZero. Slides are ready 💪

1463

Eric Brachmann @ericbrachmann.bsky.social · 19/11/2024

📢 A new learning-based approach to SfM: #ACEZero (Oral, #ECCV2024) 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. Code available. Page: nianticlabs.github.io/acezero

0494