Zhenjun Zhao

RaCo: Ranking and Covariance for Practical Learned Keypoints

Abhiram Shenoi, Philipp Lindenberger @pesarlin.bsky.social @marcpollefeys.bsky.social

tl;dr: ALIKED arch + DaD RL train+full 360rotaug. Det+covariance heads, separate ranker model. No IMC eval
#3DV2026
openreview.net/forum?id=BWt...

Marginalized Bundle Adjustment: Multi-View Camera Pose from Monocular Depth Estimates

Shengjie Zhu, Ahmed Abdelkader, Mark J. Matthews, Xiaoming Liu, Wen-Sheng Chu

tl;dr: BA for monodepth. IMC2021 results!
openreview.net/forum?id=OMT...

Understanding and Optimizing Attention-Based Sparse Matching for Diverse Local Features

Qiang Wang

tl;dr: detector matters more, not descriptor! removing nearby keypoints matters! decouple det & desc; fine-tuning detector-agnostic
eval on IMC2021

arxiv.org/abs/2602.08430

POPL-KF: A Pose-Only Geometric Representation-Based Kalman Filter for Point-Line-Based Visual-Inertial Odometry

Aiping Wang, Zhaolong Yang, Shuwen Chen, Hai Zhang

tl;dr: pose-only point and line VIO

arxiv.org/abs/2602.06425

Efficient-LVSM: Faster, Cheaper, and Better Large View Synthesis Model via Decoupled Co-Refinement Attention

Xiaosong Jia, Yihang Sun, Junqi You, Songbur Wong, Zichen Zou, Junchi Yan, Zuxuan Wu, Yu-Gang Jiang

tl;dr: decouple input view encoding from target view generation

arxiv.org/abs/2602.06478

COSMOS: Coherent Supergaussian Modeling with Spatial Priors for Sparse-View 3D Splatting

Chaeyoung Jeong, Kwangsu Kim

tl;dr: Gaussians->supergaussians

arxiv.org/abs/2602.06044

Wid3R: Wide Field-of-View 3D Reconstruction via Camera Model Conditioning

Dongki Jung, Jaehoon Choi, Adil Qureshi, Somi Jeong, Dinesh Manocha, Suyong Yeon
tl;dr: Pi-3 with (ray+radial distance) point map prediction and wide/fisheye camera tokens.
arxiv.org/abs/2602.05321

NeVStereo: A NeRF-Driven NVS-Stereo Architecture for High-Fidelity 3D Tasks

Pengcheng Chen, Yue Hu, Wenhao Li, Nicole M Gunderson, Andrew Feng, Zhenglong Sun, Peter Beerel, Eric J Seibel

tl;dr: COLMAP as initialization; ZipNeRF as backbone; DROID-SLAM as pose optimization

arxiv.org/abs/2602.05423

QuantumGS: Quantum Encoding Framework for Gaussian Splatting

Grzegorz Wilczyński, Rafał Tobiasz, Paweł Gora, Marcin Mazur, Przemysław Spurek

tl;dr: viewing directions->Bloch sphere encoding->qubit states

arxiv.org/abs/2602.05047

Please also refer our work (shameless self-promotion ;)):
bsky.app/profile/eric...

Gabor Fields: Orientation-Selective Level-of-Detail for Volume Rendering

@jcondor.bsky.social, Nicolai Hermann, Mehmet Ata Yurtsever, Piotr Didyk

tl;dr: another Gaussian + Gabor work

arxiv.org/abs/2602.05081

Wid3R: Wide Field-of-View 3D Reconstruction via Camera Model Conditioning

Dongki Jung, Jaehoon Choi, Adil Qureshi, Somi Jeong, Dinesh Manocha, Suyong Yeon

tl;dr: camera ray->pixel; spherical harmonics encodes ray directions; camera model token->model prior

arxiv.org/abs/2602.05321

VGGT-Motion: Motion-Aware Calibration-Free Monocular SLAM for Long-Range Consistency

Zhuang Xiong, Chen Zhang, Qingshan Xu, Wenbing Tao

tl;dr: optical flow->dynamic/static->submap partition; overlaped frames+loop closure->Sim(3) submap alignment

arxiv.org/abs/2602.05508

TrajVG: 3D Trajectory-Coupled Visual Geometry Learning

Xingyu Miao, Weiguang Zhao, Tao Lu, Linning Xu, Mulin Yu, Yang Long, Jiangmiao Pang, Junting Dong

tl;dr: 3D tracking->camera-coordinate trajectories; couple trajectories&point maps&relative poses

arxiv.org/abs/2602.04439

CoWTracker: Tracking by Warping instead of Correlation

Zihang Lai, Eldar Insafutdinov, Edgar Sucar, Andrea Vedaldi

tl;dr: iterative warping on current estimation,then spatio-temporal transformer refines tracks

arxiv.org/abs/2602.04877

Towards Next-Generation SLAM: A Survey on 3DGS-SLAM Focusing on Performance, Robustness, and Future Directions

Li Wang, Ruixuan Gong, Yumo Han, Lei Yang, Lu Yang, Ying Li, Bin Xu, Huaping Liu, Rong Fu

tl;dr: in title

arxiv.org/abs/2602.04251

S-MUSt3R: Sliding Multi-view 3D Reconstruction

Leonid Antsfeld, Boris Chidlovskii, Yohann Cabon, @vincentleroy.bsky.social, Jerome Revaud

tl;dr: sliding-window MUSt3R; overlapped input segments->MUSt3R->loop closure->PGO with iterative solver

arxiv.org/abs/2602.04517

Neural Predictor-Corrector: Solving Homotopy Problems with Reinforcement Learning The Homotopy paradigm, a general principle for solving challenging problems, appears across diverse domains such as robust optimization, global optimization, polynomial root-finding, and sampling. Pra...

Check out the paper for more technical details:
📄 arxiv.org/abs/2602.03086

Proud to collaborate with Jiayao Mai, Bangyan Liao, Yingping Zeng, Haoang Li, @jcivera.bsky.social, Tailin Wu, Yi Zhou, Peidong Liu 🙌

#ICLR2026 #ReinforcementLearning #Optimization #ComputerVision #MachineLearning

6/6

💡 Why it matters:

This is the first work to show that diverse homotopy solvers can be unified and learned end-to-end, opening the door for learning-based solutions across optimization, sampling, and beyond

💡 One framework
💡 Four problem domains
💡 Zero manual tuning
💡 Fast inference

5/

📊 NPC achieves consistent speedups across 4 diverse tasks:

🔹 Point Cloud Registration (GNC): 70-80% fewer iterations
🔹 Global Optimization (GH): 30-50% faster than classical methods
🔹 Polynomial Root-Finding (HC): 45% iteration reduction
🔹 Sampling (ALD): 74% fewer steps, comparable quality

4/

⚙️ How does NPC work?

At each homotopy level, our RL agent:
• Observes: current level, corrector stats, convergence velocity
• Decides: predictor step size & corrector tolerance
• Learns: to balance accuracy & efficiency via PPO

🎯 Train once on a problem class → deploy on any new instance

3/

🤔 What's the big idea?

🔍 Homotopy methods appear everywhere, from GNC in robotics to annealed Langevin in sampling, but they all rely on hand-crafted heuristics

🧠 We unify them under one framework and replace heuristics with learned policies via reinforcement learning

2/

🎉 Thrilled to share our ICLR 2026 paper:

🔹 NPC
Neural Predictor-Corrector: Solving Homotopy Problems with Reinforcement Learning

🚀 The first unified framework that reveals diverse problems share a common predictor-corrector structure

📄 Paper: arxiv.org/abs/2602.03086
💻 Code: [Coming soon]

1/

Pi-GS: Sparse-View Gaussian Splatting with Dense π^3 Initialization

Manuel Hofer, Markus Steinberger, Thomas Köhler

tl;dr: in title; π^3+PGSR

arxiv.org/abs/2602.03327

XRefine: Attention-Guided Keypoint Match Refinement

Jan Fabian Schmid, Annika Hagemann

tl;dr: pairwise kpt2subpix with self-attention and study how kpt accuracy influences the camera pose accuracy
arxiv.org/abs/2601.12530

The new CTU Rector begins their term in office with strong support for excellence. CTU has just launched a Starting Grant to attract outstanding early‑career researchers who wish to join CTU and establish their own research group. Funding: up to €160k per year for 3 years. Deadline: 30 March 2026.

tl;dr: Hutchinson’s method->Hessian matrix diagonals->second-order update; squared Hellinger distance->trust-region bounds

3DGS2-TR: Scalable Second-Order Trust-Region Method for 3D Gaussian Splatting

Roger Hsiao, Yuchen Fang, Xiangru Huang, Ruilong Li, Hesam Rabeti, Zan Gojcic, Javad Lavaei, James Demmel, Sophia Shao

arxiv.org/abs/2602.00395

Interacted Planes Reveal 3D Line Mapping

Zeran Ke, Bin Tan, Gui-Song Xia, Yujun Shen, Nan Xue

tl;dr: PlanarSplatting extension; 2D lines<->2.5D depth&normal & 2D line & 3D planar edges association->joint line-plane optimization

arxiv.org/abs/2602.01296