iiinova — where your digital self keeps evolving

David Nordström @davnords.bsky.social · 13/09/2026

New matcher unlocked. Also some VGGT-interp stuff

David Nordström @davnords.bsky.social · 21/07/2026

As an outsider, i am still curious when they are needed (i see here for RS, anything else where roma+ac is actually better than roma)?

010

David Nordström @davnords.bsky.social · 21/07/2026

My feeling is that the sparse descriptors already encode enough info to decode into 2x2 local affine. Will hopefully drop something after summer :)

110

David Nordström @davnords.bsky.social · 21/07/2026

I am inclined to agree. The dense flow is currently vastly superior to anything else. I am cooking up getting them from an already strong sparse matcher. Seems promising

220

David Nordström @davnords.bsky.social · 22/06/2026

Cool! The epicenter of the ff3d detonation

010

David Nordström @davnords.bsky.social · 18/06/2026

Malmö falafel is apparently goated

020

David Nordström @davnords.bsky.social · 18/06/2026

Lomma (LoMa) strand, bring bath trunks :). 16-18C in the ocean, mild weather, it is going to be awesome!

230

David Nordström @davnords.bsky.social · 18/06/2026

Niiiice! Sweden represent 🇸🇪

010

David Nordström @davnords.bsky.social · 17/06/2026

Too many reddit refreshes to feel proud...

000

David Nordström @davnords.bsky.social · 17/06/2026

Easy to mess up intrinsics in pose eval when you do crazy crops :) I have only tried squishing, inhereting @parskatt.bsky.social's system

010

David Nordström @davnords.bsky.social · 02/06/2026

First image matching workshop but already know it is goated

010

David Nordström @davnords.bsky.social · 02/06/2026

Heading to #CVPR26 now! You'll find me at the Image Matching Worskhop on June 4th talking about rotation invariant matching and on June 6th presenting MuM as a poster. If you are in Denver, let's meet up!

120

Reposted by David Nordström

Dmytro Mishkin @ducha-aiki.bsky.social · 29/05/2026

Image matching since 2020: 2020: @pesarlin.bsky.social SuperGlue 2023: @vincentleroy.bsky.social DUSt3R 2024: @parskatt.bsky.social RoMa 2025: @jianyuanwang.bsky.social VGGT 2026: @davnords.bsky.social"hold my beer" scales LightGlue) 2026: @jianyuanwang.bsky.social "no, hold MY beer"(scales VGGT)

0225

David Nordström @davnords.bsky.social · 15/05/2026

VGGT v2? I kinda agree Ω sounds cooler. Will adapt this notation for follow-up works going forward

000

David Nordström @davnords.bsky.social · 15/05/2026

Interesting they still have a "pointmap" loss term though skipping the head. Also glad they got rid of the horrid multi-step pass through the camera head.

010

David Nordström @davnords.bsky.social · 07/05/2026

Congratulations! Professormaxxing with 2/2 @ #ICML26

150

David Nordström @davnords.bsky.social · 04/05/2026

Not submitting to neurips but feels very likely to be the case. Last neurips this was possible and all conferences I have submitted to since

110

David Nordström @davnords.bsky.social · 02/05/2026

The AoE struggle is real :). Good luck to you all.

010

David Nordström @davnords.bsky.social · 23/04/2026

Some favorites of mine are GELU (11k cites) and Scaling Laws (8k cites)

020

David Nordström @davnords.bsky.social · 22/04/2026

Hehe I think this one is for @parskatt.bsky.social, it is from DeDoDe, right? My feeling is that this dark magic is related to gradients growing out of control with constant adding of residuals so you can regularize by sqrt(2), but not sure :)

130

David Nordström @davnords.bsky.social · 22/04/2026

Catastophic :). My girlfriend (also went to Chalmers) is furious. What do you think?

110

David Nordström @davnords.bsky.social · 20/04/2026

Classic Claude correcting you: "I just went ahead and changed the DINOv3 name to v2 as you must be confusing it".

020

David Nordström @davnords.bsky.social · 20/04/2026

My final thought is that it is not great to use a frozen encoder trained on pixel reconstruction. It works, but it can be a little wonky. I hope to get some latent multi-view objective working sometime in the future.

110

David Nordström @davnords.bsky.social · 20/04/2026

We did also try finetuning the encoder in RoMa v2 (similar to UFM) and for this use-case I was very bullish on MuM. However, we experienced training instabilities when finetuning the encoder so we dropped that.

110

David Nordström @davnords.bsky.social · 20/04/2026

Hehe it is a good point :). Not saving it for another paper (though I am playing around with a MuM v2). We tried using MuM in RoMa v2 but it did not seem to make a difference. Also, MuM representations, in contrast to DINO, are only good at the last couple of layers.

210

David Nordström @davnords.bsky.social · 18/04/2026

The days where all new 3dv pap3rs were dust3r extensions are over. All hail the new lord, inference optimization of vggt

020

Reposted by David Nordström

Dmytro Mishkin @ducha-aiki.bsky.social · 17/04/2026

It seems, that we have failed the communication about IMC26. Let's try again. The competition this year is here: kaggle.com/competitions... No prizes, but whole year leaderboard -- similar to KITTY and other academic competitions. 3D people, please retweet and share.

kaggle.com

Image Matching Challenge 2025 Ongoing

Ongoing leaderboard for Image Matching Challenge 2025.

0137

David Nordström @davnords.bsky.social · 16/04/2026

True, as mentioned in the other comment, that the activations might be quite heavy from e.g. VGGT. Some learned scene compression might work

000

David Nordström @davnords.bsky.social · 16/04/2026

Seems to be some major outage, rip

020

David Nordström @davnords.bsky.social · 16/04/2026

Interesting, that makes sense. Possibly you could train some bottleneck scene representation that forces a compression rather than storing the full VGGT activations.

010

David Nordström @davnords.bsky.social · 16/04/2026

Thanks. My thinking would be that in many cases it would be fine with an e.g. 10 second mapping stage where you run a forward pass through your multi-view transformer and thereafter you can run lightweight decoding for query images in real-time. For training you could cache this map.

110

David Nordström @davnords.bsky.social · 16/04/2026

Side note: pretty cool you managed to get a reviewer to update from 2 to 8 after a strong rebuttal. Will live on that hopium for ECCV rebuttals...

110

David Nordström @davnords.bsky.social · 16/04/2026

I guess my main curiosity is: What drove your decision to aggregate the scene information with this patch-mixing with the DUSt3R encoder rather than leveraging an encoding, like VGGT, that is natively multi-view?

100

David Nordström @davnords.bsky.social · 16/04/2026

I assume you could insert known camera poses into that representation in a similar fashion, i.e. ray encodings

100

David Nordström @davnords.bsky.social · 16/04/2026

Congratulations, cool approach of mixing patches from the reference images. Did you ever experiment with something like VGGT to encode the scene representation, possibly cache that, and then train the decoder to regress query images given that scene representation?

200

David Nordström @davnords.bsky.social · 14/04/2026

As usual, this is a collaboration with the Swedish CV special task force, i.e. @parskatt.bsky.social @fredkahl.bsky.social and @bokmangeorg.bsky.social

070

David Nordström @davnords.bsky.social · 14/04/2026

LoMa-R is our newest addition to the LoMa family. In our IMW paper at #CVPR26, we investigate rotation invariance in the sparse matching pipeline. The resulting model is robust to rotations, even matching star constellations, and achieves strong upright performance. github.com/davnords/loma

github.com

GitHub - davnords/LoMa: LoMa: Local Feature Matching Revisited

LoMa: Local Feature Matching Revisited. Contribute to davnords/LoMa development by creating an account on GitHub.

3101

Reposted by David Nordström

Georg Bökman @bokmangeorg.bsky.social · 14/04/2026

Sparse image matching is done via 1) keypoint detection in each image, 2) keypoint description, 3) matching of descriptions between images. Should rotation invariance be enforced at stage 2 or 3? Turns out both work fine! To be presented at the CVPR image matching workshop by @davnords.bsky.social

0102