Mathias Gehrig's Avatar

Mathias Gehrig

@mathiasgehrig

Computer Vision, Robotics, Machine Learning. CV/ML research engineer at Skydio. Prev. PhD in CV/Robotics at RPG in Zurich.

1,151
Followers
88
Following
149
Posts
17.11.2024
Joined
Posts Following

Latest posts by Mathias Gehrig @mathiasgehrig

That's just codex. CC does not ask for a full match, I think

22.02.2026 00:43 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

I wonder if that also holds for register tokens

10.02.2026 23:13 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

Good ol optical flow papers, not a bad one either!

08.12.2025 22:05 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Preview
The Electric Motor and the Drill - we use AI in the wrong way Power tools are better than general purpose tools for most applications, my science planning app planyourscience.com is a result of this philosophy

In which @kordinglab.bsky.social argues LLMs are more like an electric motor than a drill, and starts to build a drill for scientific research.

open.substack.com/pub/kording/...

05.12.2025 15:54 ๐Ÿ‘ 18 ๐Ÿ” 4 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

Thanks! Looking forward to trying it out :)

24.11.2025 20:29 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

Thanks! I meant table 7 for the monocular comparison. Essentially, I wonder whether this approach (with sampling n=1) is a better approach for a monocular encoder instead of the DinoV3 objective given the same compute? Right now it's not clear to me because DinoV3 was using much more compute.

24.11.2025 16:40 ๐Ÿ‘ 2 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

nice work!
I am somewhat surprised that MAE is worse than DinoV3 on the monocular dense benchmark. Do you have an intuition why this is the case?
Have you also tried sampling a single view instead of only 2 or more views? Mostly wondering if that has negative side-effects on the outcome.

24.11.2025 14:57 ๐Ÿ‘ 3 ๐Ÿ” 0 ๐Ÿ’ฌ 2 ๐Ÿ“Œ 0

Nice! Who wants to go outside in bad weather anyway

22.11.2025 02:48 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

It's unfortunate that not open sourcing the models is the best strategy against this attack

16.10.2025 10:55 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image

It's interesting that this seems so controversial yet clearly it's a correct statement.

27.09.2025 16:43 ๐Ÿ‘ 3 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image

Even Iโ€™m not sure this will end well. Graphics courtesy of Dylan Patel of SemiAnalysis.

23.09.2025 05:42 ๐Ÿ‘ 46 ๐Ÿ” 17 ๐Ÿ’ฌ 4 ๐Ÿ“Œ 1

Is sky segmentation happening?

16.09.2025 16:05 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0
Post image

Found this near Lugano. It cannot be a coincidence.

15.09.2025 09:21 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

Channel-last for everything please

02.09.2025 20:44 ๐Ÿ‘ 4 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

Before that, the EU railway system has to be unified and fixed. Want to book a train from Zurich to Barcelona? Impossible (not a joke)

26.08.2025 05:40 ๐Ÿ‘ 2 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0
Post image

excuse me sir, could you align the y-axis?

07.08.2025 18:15 ๐Ÿ‘ 3 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 1

With bad actors I mean terrorist states and organizations. They don't care. You are probably right that at least some usage in civilized parts of the world will be prevented.

01.08.2025 15:27 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

Good summary of the state of social media in 2025

31.07.2025 23:13 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

Bad actors read licenses?

31.07.2025 23:10 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

On the neuron level that makes sense to me. Neurons are often modeled as ODEs (with feedback, though)

10.07.2025 19:03 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Preview
Iterative Residual Refinement for Joint Optical Flow and Occlusion Estimation Deep learning approaches to optical flow estimation have seen rapid progress over the recent years. One common trait of many networks is that they refine an initial flow estimate either through multip...

Though, they should have mentioned this work as well: arxiv.org/abs/1904.05290

09.07.2025 10:37 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Preview
WAFT: Warping-Alone Field Transforms for Optical Flow We introduce Warping-Alone Field Transforms (WAFT), a simple and effective method for optical flow. WAFT is similar to RAFT but replaces cost volume with high-resolution warping, achieving better accu...

For optical flow nerds: Iterative image warping strikes back

arxiv.org/abs/2506.21526

09.07.2025 10:27 ๐Ÿ‘ 3 ๐Ÿ” 1 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

If someone complains you know that they read your thesis

08.07.2025 08:09 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Preview
Show Your Stripes Visualising how the climate has changed for every country across the globe

๐Ÿซ 

showyourstripes.info/c/europe/fra...

21.06.2025 10:13 ๐Ÿ‘ 14 ๐Ÿ” 4 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

Jokes aside, I hope you get better soon!

20.06.2025 16:16 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

80% dropout is still too high. You should inform your body about that

20.06.2025 15:32 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

Wow that's high quality jazz!

18.06.2025 21:00 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

I hope she also condemned Hamas in her speech as well as Iran. Because the truth is not Israel == evil but that hatred == evil. And there is plenty of hatred around that must be called out.

17.06.2025 16:05 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

Drawing the line of autonomy too early will eventually also lead to inferior systems and make you lose a war. One could argue that more autonomy is justified to stay alive.

04.06.2025 07:18 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image Post image Post image Post image

Aircraft carriers of the world
USA
UK
France
Ukraine

01.06.2025 19:19 ๐Ÿ‘ 7412 ๐Ÿ” 1591 ๐Ÿ’ฌ 49 ๐Ÿ“Œ 101