Zhenjun Zhao

🚀 Excited to share our new paper led by Zhisong and Fan (@YuishikiFan)

F²SLAM
Turning Feed-Forward Geometry into Persistent Factors for SLAM

Feed-forward 3D predictions become measurements inside dense bundle adjustment, not a one-shot correction

📉 Uncalibrated Replica ATE: 0.002 m vs 0.030 m for the best feed-forward baseline

📄 arxiv.org/abs/2610.05207
💻 github.com/ZhisongXu/F2SL… (code coming upon acceptance)

1/5

FreeView: Robust 3D Reconstruction from Uncurated Views

Junho Kim, Hyunyoung Choi, Seongwon Lee

tl;dr: select a filtering reference and reject distractor views from an uncurated collection

arxiv.org/abs/2610.33656

ProMo3D: Probing Motion Cues in Frozen 3D Foundation Models

Xiaoang Zhang, Mert Kiray, Yordanka Velikova, Manoj Biswanath, @BusamBenjamin

tl;dr: static–dynamic separation is decodable from frozen token representations

arxiv.org/abs/2610.33802

PosePlaner: Denoising Any Feed-Forward Pose Predictor from Pairwise Planar Geometry

Yihang Chen, Bin Tan, Yujun Shen, @YinghaoXu1, @NanXue7

tl;dr: rotation refinement conditioned on matched planar primitives (from PLANA3R); local solver for translation

arxiv.org/abs/2610.34845

Any3R: Feedforward 3D Reconstruction from Any Camera Type

Jingong Chen, Qingwen Zhang, @denghilbert, Chulwoo Pack, Olaf Wysocki, Boris Jutzi, Kwanghee Won, Kyle Gao

tl;dr: VGGT supports any mix of central cameras

arxiv.org/abs/2610.36150