Johan Edstedt
@parskatt.bsky.social
PhD student @ Linköping University
I like 3D vision and training neural networks.
Code: https://github.com/parskatt
Weights: https://github.com/Parskatt/storage/releases/tag/roma
I like 3D vision and training neural networks.
Code: https://github.com/parskatt
Weights: https://github.com/Parskatt/storage/releases/tag/roma
Reposted by Johan Edstedt
SPIDER: Spatial Image CorresponDence Estimator for Robust Calibration
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
November 25, 2025 at 3:11 PM
SPIDER: Spatial Image CorresponDence Estimator for Robust Calibration
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
Pixel reconstruction has recently been somewhat overshadowed by latent SSL approaches such as DINO.
However, for 3D tasks we show that a scaled and simplified version of multi-view MAE (which we call MuM) can outperform DINOv3, all while using orders of magnitude less compute!
However, for 3D tasks we show that a scaled and simplified version of multi-view MAE (which we call MuM) can outperform DINOv3, all while using orders of magnitude less compute!
We are introducing MuM, a feature encoder (ViT-L) tailored for 3D vision tasks.
TLDR; Spiritual successor to CroCo with a simpler multi-view objective and larger scale. Beats DINOv3 and CroCo v2 in RoMa, feedforward reconstruction, and rel. pose.
arxiv.org/abs/2511.17309
github.com/davnords/mum
TLDR; Spiritual successor to CroCo with a simpler multi-view objective and larger scale. Beats DINOv3 and CroCo v2 in RoMa, feedforward reconstruction, and rel. pose.
arxiv.org/abs/2511.17309
github.com/davnords/mum
November 24, 2025 at 11:13 AM
Pixel reconstruction has recently been somewhat overshadowed by latent SSL approaches such as DINO.
However, for 3D tasks we show that a scaled and simplified version of multi-view MAE (which we call MuM) can outperform DINOv3, all while using orders of magnitude less compute!
However, for 3D tasks we show that a scaled and simplified version of multi-view MAE (which we call MuM) can outperform DINOv3, all while using orders of magnitude less compute!
Put up my simple skysegmentation model on github over at github.com/Parskatt/sky...
Results are pretty crisp, but it doesn't really deal with clouds (it's literally just a linear model on top of some coarse segmentation output).
Results are pretty crisp, but it doesn't really deal with clouds (it's literally just a linear model on top of some coarse segmentation output).
November 21, 2025 at 3:11 PM
Put up my simple skysegmentation model on github over at github.com/Parskatt/sky...
Results are pretty crisp, but it doesn't really deal with clouds (it's literally just a linear model on top of some coarse segmentation output).
Results are pretty crisp, but it doesn't really deal with clouds (it's literally just a linear model on top of some coarse segmentation output).
RoMa v2 is now out! (github.com/Parskatt/rom..., arxiv.org/abs/2511.15706)
Here are the main improvements we made since RoMa:
Here are the main improvements we made since RoMa:
November 20, 2025 at 9:25 AM
RoMa v2 is now out! (github.com/Parskatt/rom..., arxiv.org/abs/2511.15706)
Here are the main improvements we made since RoMa:
Here are the main improvements we made since RoMa:
Can someone more familiar with sota diffusion tell me what's currently typically used, and does it matter at scale?
November 18, 2025 at 1:21 PM
Can someone more familiar with sota diffusion tell me what's currently typically used, and does it matter at scale?
Reviewers will be released upon acceptance of the manuscript.
November 14, 2025 at 7:13 AM
Reviewers will be released upon acceptance of the manuscript.
somewhat niche meme
November 13, 2025 at 7:08 AM
somewhat niche meme
Accepted to #3DV2026!
Radially Distorted Homographies, Revisited
Mårten Wadenbäck, Marcus Valtonen Örnhag, @parskatt.bsky.social
tl;dr: minimal solvers for one-sided/two-sided equal/two-sided independent radial distortion homography
arxiv.org/abs/2508.21190
Mårten Wadenbäck, Marcus Valtonen Örnhag, @parskatt.bsky.social
tl;dr: minimal solvers for one-sided/two-sided equal/two-sided independent radial distortion homography
arxiv.org/abs/2508.21190
November 7, 2025 at 7:47 AM
Accepted to #3DV2026!
I don't want to submit to any conference that would accept my paper.
November 6, 2025 at 12:51 PM
I don't want to submit to any conference that would accept my paper.
CVPR compute form is so incredibly dumb.
November 6, 2025 at 8:16 AM
CVPR compute form is so incredibly dumb.
Did everyone get this?
November 4, 2025 at 8:41 AM
Did everyone get this?
why is this so true
AGI is when models write good multi-dim torch indexing code.
October 29, 2025 at 9:16 AM
why is this so true
I just put `MultiScaleDeformableAttention` wheels on pypi under `ms-deform-attn`.
Should make stuff involving mask2former a bit smoother.
Should make stuff involving mask2former a bit smoother.
October 21, 2025 at 7:31 PM
I just put `MultiScaleDeformableAttention` wheels on pypi under `ms-deform-attn`.
Should make stuff involving mask2former a bit smoother.
Should make stuff involving mask2former a bit smoother.
Gonna start using this
October 21, 2025 at 7:14 AM
Gonna start using this
Pro-tip: Set the loss range in your wandb plot unreasonably low to motivate your models to try harder.
October 20, 2025 at 2:09 PM
Pro-tip: Set the loss range in your wandb plot unreasonably low to motivate your models to try harder.
AGI is when models write good multi-dim torch indexing code.
October 16, 2025 at 8:12 PM
AGI is when models write good multi-dim torch indexing code.
Cursor tab likes to sprinkle in some "delete 20 lines of code" seemingly at random.
I like that it adds a bit of excitement to the coding experience.
I like that it adds a bit of excitement to the coding experience.
October 16, 2025 at 5:30 PM
Cursor tab likes to sprinkle in some "delete 20 lines of code" seemingly at random.
I like that it adds a bit of excitement to the coding experience.
I like that it adds a bit of excitement to the coding experience.
Spooky features
Pro tip: For good Halloween vibes, use non-normalized RoPE on images larger than your training resolution and larger than the composite period of some of the RoPE-rotations. You might get scary ghost structures in your features.
October 16, 2025 at 3:10 PM
Spooky features
Sweden is a Monarchy that functions like a Democracy.
Some other countries do it the other way round.
Must say I prefer the former.
Some other countries do it the other way round.
Must say I prefer the former.
October 16, 2025 at 10:52 AM
Sweden is a Monarchy that functions like a Democracy.
Some other countries do it the other way round.
Must say I prefer the former.
Some other countries do it the other way round.
Must say I prefer the former.