Sign in

Karim Knaebel

@karimknaebel.bsky.social
1.4K followers 691 following 86 posts

PhD @RWTH.bsky.social | 3D Computer Vision previously Karim Abou Zeid 🔗 karimknaebel.com

PostsRepliesMedia
Karim Knaebel @karimknaebel.bsky.social · 02/06/2026
thank you for sharing :)
020
Reposted by Karim Knaebel
Zhenjun Zhao @ericzzj.bsky.social · 01/06/2026
SurGe: Improved Surface Geometry in Point Maps @karimknaebel.bsky.social, Gonzalo Martin Garcia, Christian Schmidt, @ifradlin.bsky.social, Lucas Nunes, @dcdegeus.bsky.social, Bastian Leibe arxiv.org/abs/2605.31577
253
Karim Knaebel @karimknaebel.bsky.social · 14/05/2026
I think it’s important to have a policy, and actually enforce it in unambiguous cases. This is separate from finding these cases. You won’t find all of it, but at least when such a case is revealed by eg auto flagging bogus references, user reports, or coincidence, there is a policy in place.
010
Karim Knaebel @karimknaebel.bsky.social · 13/05/2026
Bravo! arXiv soon stricter than major ML conferences? 😁
110
Karim Knaebel @karimknaebel.bsky.social · 04/05/2026
kk thx! yes, I also think so, but then I couldn't actually find this explicitly stated anywhere.
110
Karim Knaebel @karimknaebel.bsky.social · 04/05/2026
@neuripsconf.bsky.social will we still be able to change the abstract and title after the abstract deadline?
100
Karim Knaebel @karimknaebel.bsky.social · 29/04/2026
Is this breakfast? 😅
010
Karim Knaebel @karimknaebel.bsky.social · 07/04/2026
I like this git-diff-style figure a lot.
120
Karim Knaebel @karimknaebel.bsky.social · 14/03/2026
Oh no, I love to do that. Counting my days 🥲
120
Karim Knaebel @karimknaebel.bsky.social · 23/02/2026
Do you also add a short cooldown phase after the constant part? I believe for self-supervised representation learning you don’t want to cool down, but for downstream tasks it usually gives a big boost.
010
Karim Knaebel @karimknaebel.bsky.social · 08/02/2026
were cvpr reviews ever public since this motion?
110
Karim Knaebel @karimknaebel.bsky.social · 30/01/2026
I ran the first frame through an (undertrained) internal monocular model. When you zoom in, you can see the tree and street for scale. The clouds still seem too small though; only about 42x the size of the tree in the foreground.
040
Karim Knaebel @karimknaebel.bsky.social · 12/01/2026
How is it broken? (Genuine question)
100
Karim Knaebel @karimknaebel.bsky.social · 10/11/2025
Those are awesome, thanks for the inspiration
030
Karim Knaebel @karimknaebel.bsky.social · 10/10/2025
Congrats!!
010
Karim Knaebel @karimknaebel.bsky.social · 30/09/2025
So you uv sync -p 3.13 the first time (if you care about the exact version), and then you never need to pass it again. The existing venv “soft pins” its python version until you explicitly request a different version, either via CLI or updated constraints of the python version.
030
Karim Knaebel @karimknaebel.bsky.social · 30/09/2025
I use it all the time, super useful :) It won’t pin, just makes sure your venv uses this python version if it doesn’t already or doesn’t exist yet.
100
Karim Knaebel @karimknaebel.bsky.social · 29/09/2025
What’s the reason for pinning in general? I somehow never found a use case for it. If I want a specific version from that range, I provide it in uv sync on demand.
100
Karim Knaebel @karimknaebel.bsky.social · 21/09/2025
Very good read, thanks for writing this up :)
010
Karim Knaebel @karimknaebel.bsky.social · 10/09/2025
Does it still happen when you compile before DDP? I always use DDP as the outermost wrapper, there is too much hacky stuff going on in its forward and backward for my liking.
100
Karim Knaebel @karimknaebel.bsky.social · 25/07/2025
The normal metrics in Tab. 1 also don’t match. I checked only for DSINE, and here the actual metrics are considerably better than in Tab 1. Also missing some sota normal methods. Tab 3 is only a small subset of common benchmarks, and on 1/2 of the datasets it’s worse than both VGGT and MoGe.
120
Karim Knaebel @karimknaebel.bsky.social · 21/07/2025
Same, I don’t want to use them anymore now, and if I “have to”, then I’m writing them as triple dashes ---. At least semicolons are still safe 👀
000
Reposted by Karim Knaebel
International Conference on 3D Vision @3dvconf.bsky.social · 01/07/2025
Understanding and reconstructing the 3D world are at the heart of computer vision and graphics. At #CVPR2025, we’ve seen many exciting works in 3D vision. If you're pushing the boundaries, please consider submitting your work to #3DV2026 in Vancouver! (Deadline: Aug. 18, 2025)
1177
Karim Knaebel @karimknaebel.bsky.social · 17/06/2025
In my mind, I often associate instance seg with fancy depth clustering.
020
Karim Knaebel @karimknaebel.bsky.social · 13/04/2025
I like the look. DINOv2, DUSt3R, CroCo v2, and GB are misspelled if you want to fix those.
110
Karim Knaebel @karimknaebel.bsky.social · 10/04/2025
I haven’t experienced that yet, but I’m not sure if I’ve tried on that many files. Which one did you try: `dua a` (or just `dua`) will block and finally print, which can take a long time. `dua i` starts an interactive TUI and loads everything in the background while you can already start inspecting.
000
Karim Knaebel @karimknaebel.bsky.social · 10/04/2025
Looks good. I haven’t seen it before. Can also recommend this one: github.com/Byron/dua-cli
github.com
GitHub - Byron/dua-cli: View disk space usage and delete unwanted data, fast.
View disk space usage and delete unwanted data, fast. - Byron/dua-cli
210
Reposted by Karim Knaebel
David Picard @davidpicard.eurosky.social · 21/03/2025
🔥🔥🔥 CV Folks, I have some news! We're organizing a 1-day meeting in center Paris on June 6th before CVPR called CVPR@Paris (similar as NeurIPS@Paris) 🥐🍾🥖🍷 Registration is open (it's free) with priority given to authors of accepted papers: cvprinparis.github.io/CVPR2025InPa... Big 🧵👇 with details!
713652
Karim Knaebel @karimknaebel.bsky.social · 04/04/2025
Ah yes, you’re right, thanks for the reference! It’s really just a line in space without explicitly specifying a point on the line as its origin.
000
Karim Knaebel @karimknaebel.bsky.social · 04/04/2025
The origin is not lost though, it’s just encoded differently. In the end it’s still 6 dims, 3 for the direction and 3 for the origin. A bit unclear to me though what benefit this origin encoding brings.
100
Karim Knaebel @karimknaebel.bsky.social · 04/04/2025
Their proposed architecture only makes sense for a dense parameterization. I don’t think they ablate with other dense representations though.
010
Karim Knaebel @karimknaebel.bsky.social · 04/04/2025
Correction: the origin is already encoded in the Plücker ray. I meant that DiffusionSfM encodes /both/ origin and endpoint.
010
Karim Knaebel @karimknaebel.bsky.social · 04/04/2025
I think the first time I saw Plücker rays was in the Camera as Rays paper arxiv.org/abs/2402.14817. The same authors also recently released DiffusionSfM, where they indeed keep the origins o as well.
arxiv.org
Cameras as Rays: Pose Estimation via Ray Diffusion
Estimating camera poses is a fundamental task for 3D reconstruction and remains challenging given sparsely sampled views (<10). In contrast to existing approaches that pursue top-down prediction of gl...
210
Karim Knaebel @karimknaebel.bsky.social · 30/03/2025
Agree, einops or shape in comments is a must. Also, makes it much easier for others to read and understand the code without executing it.
000
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
Maybe last one, could you try VGGT single-view on that image for direct comparison? 😁
100
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
Very crisp!
100
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
True. Does MoGe do a better job at segmenting the sky here?
100
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
Hmm yes, not sure if the demo is bugged or if it didn’t correctly segment the sky 🤔 Do you have access to GPUs to test this locally?
100
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
That's what MoGe is doing
120
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
Feels more native than a separate mask prediction as with MoGe. Then again, could turn out to be better not to mix up the concepts of confidence and undefinedness even though they seem related.
110
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
Ah yes, now I get it. Detecting sky is usually not part of the benchmarks.
000
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
There is no supervision for specific confidence values though, if that’s what you meant.
100
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
The depth loss is scaled by the confidence and the sum of confidences is added as a regularization term, otherwise the model just predicts all-zero confidence and gets a perfect loss.
100
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
The cleanest way would probably be to force the model to predict zero confidence for sky pixels, but still mask both the sky-pixel confidence and depth values from the other losses.
110
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
Regressing huge values close to inf is hard/impossible to do with these model architectures and will make training unstable. I think masking inf values during training and eval makes sense. For inference w/o GT either a separate segmentation model, or predicting a mask as in MoGe is perfectly fine.
210
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
They use an of-the-shelf sky segmentation method.
010
Karim Knaebel @karimknaebel.bsky.social · 20/03/2025
Benchmarks also ignore sky pixels; otherwise, what would you say should be the correct target for these in a benchmark?
110
Karim Knaebel @karimknaebel.bsky.social · 19/03/2025
How does it look with sky pixels masked out in the output? I think I read that sky pixels are undefined outputs in VGGT, because they’re excluded from the loss.
200
Karim Knaebel @karimknaebel.bsky.social · 17/03/2025
Awesome work, congrats!
010
Reposted by Karim Knaebel
Jianyuan Wang @jianyuanwang.bsky.social · 17/03/2025
Introducing VGGT (CVPR'25), a feedforward Transformer that directly infers all key 3D attributes from one, a few, or hundreds of images, in seconds! Project Page: vgg-t.github.io Code & Weights: github.com/facebookrese...
34614