Michaël Ramamonjisoa

Researcher @MetaAI (FAIR) Paris | PhD in 3D Computer Vision from @ImagineEnpc ex intern @RealityLabs and @NianticLabs

[Michaël Ramamonjisoa @MichaelRamamon]

DINOv3 is out! Super proud of our team's contribution to the computer vision community. Check out this great summary by @BaldassarreFe to understand dig deep in our features and how we got there!
Now let's focus on applications enabled by DINOv3👇

[Federico Baldassarre @BaldassarreFe]

Say hello to DINOv3 🦖🦖🦖

A major release that raises the bar of self-supervised vision foundation models.
With stunning high-resolution dense features, it’s a game-changer for vision tasks!

We scaled model size and training data, but here's what makes it special 👇

[Elouan Gardès @elouangardes]

What if the metadata from acquisition was enough? 👀

Sometimes the most useful labels aren't labels at all.

Introducing📄 Who Needs Labels? Adapting Vision Foundation Models With the Metadata You Already Have

[Nicolas DUFOUR @nico_dufour]

🚀 DinoV3 just became the new go-to backbone for geoloc!
It outperforms CLIP-like models (SigLip2, finetuned StreetCLIP)… and that’s shocking 🤯
Why? CLIP models have an innate advantage — they literally learn place names + images. DinoV3 doesn’t.

[Rui Li @leedaray]

Another major update of the "awesome-dust3r" (https://t.co/Iwx79viNOY) paper list.

There are more VGG-T follow-ups and some interesting correlations, e.g., VGGT-Long/LONG3R, STream3R/Streaming 4D-VGGT.

Let's see what happens for visual geometry in the DINO-v3 era :)

[Michaël Ramamonjisoa @MichaelRamamon]

@chrisoffner3d That's a good experiment indeed! What I was referring to above was the heavily supervised (geometrically with RGB-D samples) paradigm that metric depth estimators explore, vs stronger backbones but "light" depth supervision, like us and Marigold

[Michaël Ramamonjisoa @MichaelRamamon]

That would be awesome! One thing, I wonder how much depth data is actually needed on top of DINOv3 for metric depth. The latest works use sooo much data *and* fine-tune the backbone, it's hard what is used to specialize the representation and what goes into decoding it into depth

[Chris Offner @chrisoffner3d]

I want to see DINOv3 with Depth Pro for metric depth.

[Chris Offner @chrisoffner3d]

SigLIP (VLMs) and DINO are two competing paradigms for image encoders.

My intuition is that joint vision-language modeling works great for semantic problems but may be too coarse for geometry problems like SfM or SLAM.

Most animals navigate 3D space perfectly without language.

[Chris Offner @chrisoffner3d]

@ducha_aiki @oriane_simeoni RoMa + DINOv3 = RoMAv2
Depth Anything v2 + DINOv3 = Depth Anything v3
Depth Pro + DINOv3 = Depth Ultra

[Michaël Ramamonjisoa @MichaelRamamon]

Well, I'm sure there will be ways to optimize it! Maybe using our smaller backbones would be a start 😉

[Dmytro Mishkin 🇺🇦 @ducha_aiki]

@chrisoffner3d @oriane_simeoni Please, UFM-like speed for RoMav2, not original one..,

[Dmytro Mishkin 🇺🇦 @ducha_aiki]

@chrisoffner3d @oriane_simeoni Please, UFM-like speed for RoMav2, not original one..,

[Chris Offner @chrisoffner3d]

I want to see DINOv3 with Depth Pro for metric depth.

[Federico Baldassarre @BaldassarreFe]

For more DINOv3 evaluations, check out this thread by @MichaelRamamon! He's the expert in 3D tasks (depth, point matching) and ConvNeXt distillation. Congrats to @MarcSzafraniec and Seungeun for their work on detection and segmentation too!

[Jianyuan @jianyuan_wang]

It’s been a real pleasure playing with DINOv3 and training a new VGGT with it. 🚀 I believe its potential goes far beyond what current benchmarks can reveal. During training, you can feel the “character” of this model: smart, delicate, and surprisingly adaptable.

It’s not just about hitting SOTA numbers. It’s how DINOv3 handles complex data, uncovers structure, and stays rock-solid even in tough training regimes. I’m truly excited to see how the vision community will take this further! 🌟

Massive thanks to @MichaelRamamon, @maxseitzer, Cijo Jose, @_claireroberts, @monsieurlabatut, @p_bojanowski, and the whole team for the fantastic collaboration and incredible effort. I’m very happy to have contributed, and grateful to everyone for making this possible!

And on a personal note, I actually love the design choice of switching back to patch size 16 🤣 No more painful upsampling!

[Michaël Ramamonjisoa @MichaelRamamon]

I cannot state how happy and proud I am about the work our team has put to make DINOv3 a reality - one that could enable so many cool research and applications. If you also think good and fast visual features could benefit your application - in particular 3D ones 😊 - let's chat!

[Michaël Ramamonjisoa @MichaelRamamon]

DINOv3 is out! Super proud of our team's contribution to the computer vision community. Check out this great summary by @BaldassarreFe to understand dig deep in our features and how we got there!
Now let's focus on applications enabled by DINOv3👇

[Federico Baldassarre @BaldassarreFe]

Say hello to DINOv3 🦖🦖🦖

A major release that raises the bar of self-supervised vision foundation models.
With stunning high-resolution dense features, it’s a game-changer for vision tasks!

We scaled model size and training data, but here's what makes it special 👇

[Michaël Ramamonjisoa @MichaelRamamon]

@BaldassarreFe showed we can track masks through videos with no training! Looking at this from a 3D person perspective, you'd think you could do better feature matching right? You'd be right! I loved how I could just plug DINOv3 into Probe3D and immediately get great results!

[Piotr Bojanowski @p_bojanowski]

🔥 The DINO team is looking for a PostDoc! 🔥
If you are about to graduate, and want to be part of what’s next for SSL, don’t hesitate to reach out!
Link to job offer : https://t.co/jaUBImQg4g