3D Computer Vision at Apple. I like 3D vision and training neural networks.
- Tweets 2,626
- Following 241
- Followers 1,299
- Likes 3,116
> "we create a large-scale 3D dataset with over 1M images"
> look inside
> it's 4 slightly different rooms with 200k images each
8
9
237
22,090
The 3D vision community really hates the bitter lesson.
Dust3r is what you get when you take the lesson seriously.
10
11
193
28,145
Famously the tractor was created by predicting all recorded horse movements from the beginning of history, and then posttrained on horsing around satisfactorily.
2
101
6,432
👀 getting hyped 👀
Can't wait for Noah Snavely (@Jimantha)'s keynote talk: MegaScenes: Reconstructing All the World's Landmarks at our #USM3D workshop at @CVPR on June 17th!
ALT Keynote Speaker Annoncement for USM3D Workshop at on June 17th CVPR 2024 1st Workshop on Urban Scene Modeling: Where Vision Meets Photogrammetry and Graphics more info at usm3d.github.io Noah Snavely is the Professor of Computer Science at Cornell Tech interested in computer vision and computer graphics, and a member of the Cornell Graphics and Vision Group. He also work at Google Research in NYC. His research interests are in computer vision and graphics, in particular in 3D understanding and depiction of scenes from images. He is an ACM Fellow, and the recipient of a PECASE, a Microsoft New Faculty Fellowship, an Alfred P. Sloan Fellowship, and the SIGGRAPH Significant New Researcher Award.
2
12
98
18,176
DeDoDe is now on arxiv! 🥳
arxiv.org/abs/2308.08479
Say hi to DeDoDe 🎶!
DeDoDe is a keypoint detector trained to detect 3D tracks. The reverse, DeDoDe, is a descriptor that matches the tracks.
DeDoDe and DeDoDe are simple to train, and show great performance 📈
Code: github.com/parskatt/dedode
3
23
89
9,590
I dont understand why #CVPR2024 put all related papers in the same session. As an author it means I can't go to the posters I'm actually interested in.
5
6
93
17,244
As pdf: github.com/Parskatt/storage/…
As keynote:
github.com/Parskatt/storage/…
2
5
56
5,336
Rejected from #ECCV24 by AC with motivation that we can't add experiments (eval, requested by reviewers, not improvement of method) in the rebuttal. Pretty frustrating.
3
52
13,304
Say hi to DeDoDe 🎶!
DeDoDe is a keypoint detector trained to detect 3D tracks. The reverse, DeDoDe, is a descriptor that matches the tracks.
DeDoDe and DeDoDe are simple to train, and show great performance 📈
Code: github.com/parskatt/dedode
7
12
50
15,185
I feel like the reason people do incremental work is that it's way easier to get accepted. Would be nice if we could tip the balance a little bit towards novelty (and novelty doesn't mean complicated).
4
2
51
5,516
Objaverse is the most important 3D vision paper in last 5 years, if google made street view data accessible that would be the most important in 10 years.
4
3
40
7,835
DL3DV looks good :) @LuLing26466911 thanks for making the colmap caches available!
1
5
42
5,929
And today I'm presenting this work at #CVPR2025!
🗓️ Date: 16:00-18:00, Fri, Jun 13 (Today)
📍Place: Poster #115 in Session 2 (ExHall D)
💻 Code: github.com/ericssonresearch/…
ColabSfM: Collaborative Structure-from-Motion by Point Cloud Registration
@Parskatt , André Mateus, Alberto Jaenal
tl;dr: in title, learning to register SfM point clouds.
arxiv.org/abs/2503.17093
2
43
3,328
Image matching is pretty great to work on because there's like 5 people working on it, so any idea you come up with is probably new.
6
40
9,196
Yushan's work GMSF: Global Matching Scene Flow is accepted to NeurIPS 2023!🥳
We propose a simple but powerful approach to scene flow estimation through global matching that achieves state-of-the-art performance.
Paper: arxiv.org/abs/2305.17432
Code: github.com/ZhangYushan3/GMSF
9
39
3,602
DeDoDe v2 coming soon with some improvements to the detector! Colab with @BokmanGeorg and @zhenjun_zhao
3
7
38
5,793
Counterpoint: Even without further significant breakthroughs in capabilities, there will be significant value in getting an edge in efficiency/perf due to the size of the market.
1
38
6,543
No better feeling than dataset providing pose as 4x4 matrix in documented coordinate system.
3
2
35
2,177
DeDoDe now in kornia 😀
0.7.2 is out!
- Added DeDoDe features (thanks @Parskatt )
- LightGlue models, available nowhere else - DeDoDe (B/G), KeyNet-HardNet
- KMeans implementation
- New augmentations: RandomGaussianIllumination, RandomLinearIllumination, RandomLinearCorner
1/2
github.com/kornia/kornia/rel…
1
4
33
3,485
At least in Sweden it can also mean that the place is too local, only regulars go there and rate the place highly.
Typical example: local pizzeria.
33
1,604
There are seriously undervalued parts of 3D vision. Top among these are auto-calibration.
2
2
33
3,678
Checkout my "50% done" PhD seminar, where I talk about my recent works in 3D Reconstruction using Neural Networks!
There are new things in there! Watch to the end!
If you have questions, ask them in this thread :)
Link: nextcloud.liu.se/s/fqj99km9S…
2
2
28
5,434
The real problem with COLMAP is that it's impossible to ctrl+F stuff. There's always 3 layers of abstraction for everything. Makes it so hard when you don't have 5 years of experience with it.
2
24
2,960
Choose your imports carefully lol. Maybe I should change name to RoMatch... Or they should change to RotMan
RoMa: an easy-to-to-use, stable and efficient library to deal with rotations and spatial transformations in PyTorch.
Read all about this PyTorch Ecosystem Tool in our latest Medium post ⚡hubs.la/Q02hFsVf0
3
1
24
4,091
This is being slept on. MVS + Splats
RPBG: Towards Robust Neural Point-based Graphics in the Wild
Qingtian Zhu, Zizhuang Wei, Zhongtian Zheng, Yifan Zhan, Zhuyu Yao, Jiawang Zhang, Kejian Wu, Yinqiang Zheng
tl;dr: image restoration->reform the CNN-based neural renderer from NPBG
arxiv.org/pdf/2405.05663
4
1
22
9,743
On Thursday I'll give a talk on image matching over in Prague. Hope to meet you there :)
1
4
20
2,884
>find bug in training code
>fix implementation and rerun experiment
>the performance is now worse
2
20
1,158
We need to make some simple 3D geometric benchmarks so 2D people can evaluate their new shiny backbones. Annoying to see the 15 benchmark eval where it's all 2D classification.
1
20
2,001
vkitti2 flow tracks the rotation of the wheels, might be somewhat challenging to match lol
3
20
2,194
This is probably because it converts images to be square, in the preprocessing step. Try it with a cropped image instead and see if it improves.
19
Can someone pls make better MVS than patchmatch, I'm begging you. Otherwise I'm making it this autumn.
6
18
4,041
I'll be presenting this paper today at 13:00 at the Image Matching Workshop at #CVPR2025!
🗓️ Date: 13:00, Wed, Jun 11 (Today)
📍Room: 108
📖 Slides: github.com/Parskatt/storage/…
Less Biased Noise Scale Estimation for Threshold-Robust RANSAC
@Parskatt
tl;dr: estimate RANSAC th from data, almost optimal
P.S."χ2 model holds pretty well for correspondences, which is lucky,as our life would otherwise be significantly more difficult" arxiv.org/abs/2503.13433
2
2
20
3,070
Coordinate system tier list based on how easy they are to make with your hand. Blender S-tier, COLMAP/OpenCV F-tier.
2
1
17
1,777
Using COLMAP partial reconstructions beyond 0 is like going to the second page of google search results.
1
14
830
COLMAP undistort randomly removed 2px from images.
I removed the output and reran the exact same script.
It's working now.
I'm scared.
3
15
2,126
writing a new paper and found that I didn't do a single self-cite. not sure if I should be proud.
5
14
1,786
Car goes on a nice highway, public transport consists of 3 different buses with 20 stops.
Also you're lucky if they show up at all.
1
13
2,864
My mildy controversial take is that keypoint detectors aren't, and probably shouldn't be, scale invariant.
5
13
2,062
I agree for senior PhDs who already has several papers published. I don't agree for new students.
Your first paper wont be your best resesrch (usually), and you need practice through quantity.
2
13
522
Wake up babe, new matching dataset just dropped
Our DL3DV-10K dataset paper has been accpeted by #CVPR2024🎉! It provides scene-level videos at 4K resolution, RGB-images, camera pose, and point coulds. The DL3DV-3K is currently available and more versions come soon. Feel free to check our project page: dl3dv-10k.github.io/DL3DV-10…
1
13
1,497
Seen multiple labs seemingly very proud that they don't have any GPUs. And these are not poor labs.
2
13
2,458
Both nerfs and gsplats follow bitter lesson in the sense that they are simple and scale well with data.
1
13
1,758
Quaternions should never be exposed outside of library, happy that pycolmap banned it. Coordinate systems are difficult enough as is.
4
11
2,359
Feature matching is a crucial component of structure-from-motion (SfM) that aims to match features. While traditional handcrafted methods~\cite[sift] are traditionally used, recently learning-based methods~\cite[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16] are becoming increasingly po
12
244
Also shoutout to tf32 for ruining precision in my experiments due to deps enabling it on matmul by default. I really hate tf32.
1
12
1,790
The method of combining rocks into a henge is simple, and while a simple method is not grounds for rejection, the method is not general.
If the authors put some henges in Sudan and Thailand I might reconsider.
11
620
It might be nice to cite FlyingThings paperswithcode.com/dataset/f…
11
636