Linfei Pan

@CVPR Same question here: what should be put in the paper ID part exactly in this case?

Our #CVPR2026 Program Chairs have just pushed out a follow-up email with clarifications. Check your email (and spam folder).

It is presented at CVPR poster session 1 at stand 81 as a highlight paper. Look forward to feedbacks from the community!

@mazy1998 @ChrisAtKIRI @don2qian @visionviktor @mapo1 thanks!

@LinfeiPan @ChrisAtKIRI @don2qian @visionviktor @mapo1 Great work!

📢 Introducing GenSfM – a generic Structure-from-Motion pipeline built on COLMAP, with smooth, invertible, and universal calibration for diverse cameras — from pinhole to catadioptric! Thanks @don2qian @visionviktor @mapo1 for the great collaboration. 🔗 ivonne320.github.io/yihan.home/Gen…

🌺 Excited to announce the 1st CroCoDL Workshop on Large-Scale Cross-Device Localization at #ICCV2025! Tackling real-world localization across smartphones, AR/VR, and robots with invited talks, paper track & competition on a novel dataset.   🔗 localizoo.com/workshop 🧵 1/2

The Call for Papers is now open for our #RSS2025 workshop: Unifying Visual SLAM: From Fragmented Datasets to Scalable, Real-World Solutions If you're working on #SLAM or #localization, we’d love to see your contributions! linkedin.com/posts/alejandr…

@ducha_aiki Hi, are you using Windows? Somehow windows can give this strange bug and we still haven’t figured out the reason😥. You can try limiting the maximum track number to potentially solve it (setting TrackEstablishment.max_num_tracks < 50k or so)

I was unhappy with drift in colmap reconstruction and decided to use glomap instead. Look at this beautiful camera cube...

@pablovelagomez1 @majti89 @mapo1 @pesarlin thanks! Will look into it

@LinfeiPan @majti89 @mapo1 @pesarlin I actually already did! Made an issue about it here github.com/colmap/glomap/…

Recently, I've been working a lot on Gaussian Splatting. I built a repo for camera pose estimation using @rerundotio, @Gradio, and @prefix_dev that merges GLOMAP with hloc! 🧵More info below

@pablovelagomez1 @majti89 @mapo1 @pesarlin Hi, thanks for the effort. Do you mind sharing some of the mentioned failure cases of GLOMAP? Just for us to improve the library

I wanted to thank @LinfeiPan @majti89 @mapo1 @pesarlin for the amazing open-source repos of GLOMAP and Hloc. You can also find the original repos here github.com/cvg/Hierarchic… github.com/colmap/glomap

🚨 The CVG group at ETH Zurich is hiring! We are looking for PhD & PostDoc in egocentric vision and video understanding. 🔗 Apply by sending an e-mail to cvgresearchgroup@gmail.com and become part of our amazing team! @cvg_ethz @mapo1

@pablovelagomez1 @pesarlin @behnam_iut @janusch_patas No problem. I guess the gain for the global positioning would be similar to bundle adjustment but I would need to perform some tests to make an accurate conclusion there 😀

@LinfeiPan @pesarlin @behnam_iut @janusch_patas Got it! Thank you for the response. I'm not sure it matters that much considering how fast glomap already is 🙂

Excited to share my latest project! I've successfully ported ALikeD and LightGlue to C++/LibTorch, achieving a notable speedup. Now, I’m aiming even higher, i.e. building an entire SfM pipeline completely based on LibTorch, with the goal of running everything on the GPU for maximum speed (also prunning ceres, eigen, etc.) The big question: what's the best way forward? One option is to port Glomap and integrate it, but I’m also exploring whether there are better alternatives out there. I’m eager to hear your thoughts - what approaches or frameworks would you recommend to achieve a fast, fully GPU-based SfM pipeline? Advice and insights are highly welcome! Let’s discuss. 😊

@pesarlin @pablovelagomez1 @behnam_iut @janusch_patas Yes, GLOMAP does not support GPU acceleration yet. The reason for re-implementing BA was the difference in the base classes. Global positioning is slightly different as it has many extra parameters. I guess both can potentially benefit from GPU.

@pablovelagomez1 @behnam_iut @janusch_patas AFAIK No since GLOMAP reimplements the BA function, but it shouldn’t really need it @LinfeiPan

@ducha_aiki @majti89 @mapo1 Thanks for sharing our work!

Gravity-aligned Rotation Averaging with Circular Regression @LinfeiPan @majti89 @mapo1 tl;dr: in title, also already in Glomap ecva.net/papers/eccv_20… #ECCV2024

@ducha_aiki @majti89 @mapo1 Thanks for sharing our work!

Global Structure-from-Motion Revisited @LinfeiPan @majti89 @mapo1 J. Schönberger tl;dr: translation averaging is bad for global SfM health. Jointly optimize camera pos+pointcloud instead. Great results. Eval on #IMC2023 train set (? but fine). #ECCV2024 1/ arxiv.org/abs/2407.20219

@foxofdev @majti89 @mapo1 Thanks for your interest. We performed experiments on IMC test+val (without resizing) as in VGGSFM, and we obtained around 70 AUC scores at 3deg. So, GLOMAP performs well on traditional datasets, but might suffer similarly on textureless images as COLMAP. See limitation in paper

@majti89 @mapo1 @LinfeiPan Is it better than VGGSfM from Facebook?

🚀 Ready to take 3D reconstruction to the next level? Whether you're working on NeRF or 3DGS, our new method, GLOMAP, is here to impress! 🌟 It's faster and more accurate than COLMAP on several datasets. 🌐 Website: lpanaf.github.io/eccv24_glomap/ @mapo1, @LinfeiPan, J. Schönberger

@Coding_Black @majti89 @mapo1 Thanks! It seems that camera pose in DL3DV are obtained from SFM pipelines (COLMAP?) If so, then I am not sure to what extent our work can be directly benchmarked. However, I think it should be quite nice to use GLOMAP there to efficiently retrieve camera poses

@majti89 @mapo1 @LinfeiPan @LinfeiPan Great work! Do you think the DL3DV dataset could help show the effectiveness of GLOMAP?

@jon_barron Thanks for sharing our work!

lpanaf.github.io/eccv24_glomap/ seems really good? "In terms of accuracy and robustness, we achieve results on-par or superior to COLMAP, the most widely used incremental SfM, while being orders of magnitude faster."

@eralagee @majti89 @mapo1 Thanks for your interest! I don’t have concrete experiments to support this, but my guess is yes. Not sure to what extent this applies though

@majti89 @mapo1 @LinfeiPan Very cool work! Can we have a coarse initialization of pose, i.e., pose from RTK, to further reduce the time consuming?

@BokmanGeorg @majti89 @mapo1 Exactly. With final ba, the estimation is further refined

@BokmanGeorg @majti89 @mapo1 Thanks for your interests. We have only tried initialising with uniform pos in our experiments, and they all work well. But my guess would be that with a different initialisations, the convergence rate might be further improved