Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Image-to-Point Cloud Registration Made Easy wit...

Sponsored · Your Podcast. Everywhere. Effortlessly. Share. Educate. Inspire. Entertain. You do you. We'll handle the rest. →
Avatar for Reon Tabata Reon Tabata
November 05, 2026

Image-to-Point Cloud Registration Made Easy with Rectified Flow-based LiDAR Upsampling

IEEE/RSJ International Conference on Intelligent Robots and Systems 2026

Avatar for Reon Tabata

Reon Tabata

November 05, 2026

Other Decks in Research

Transcript

  1. Image -to -Point Cloud Registration Made Easy with Rectified Flow

    -based LiDAR Upsampling Reon Tabata*1*2, Kenji Koide*2, Shuji Oishi*2, Masashi Yokozuka *2, Taku Okawara *2, Aoki Takanose *2, Jun Miura*1 *1 Toyohashi University of Technology, Japan *2 National Institute of Advanced Industrial Science and Technology (AIST), Ja pan
  2. Overview • An image -to -point cloud registration method that

    • Rectified Flow -based intensity upsampling • Outperforms baselines; with pre -trained image feature matching no image -point cloud pairs or ground LiDAR Intensity Image Upsampling Input Image treats LiDAR as an imaging sensor Generated Image -truth poses needed for training Pre -trained Image Feature Matching Matching with LightGlue 2
  3. Objective Goal Image -to -point cloud registration with easily collected

    training data while achieving high accuracy and strong generalization Approach Image -to -image registration can leverage powerful pre -trained matchers, even across different modalities. 1 Treat LiDAR as an imaging sensor via intensity upsampling Estimate relative pose using image -to -image registration 2 Self -supervised pre -training on readily available images Fine -tuning with a small amount of LiDAR data 5
  4. Method [2/4] Training Conditional Rectified Flow Pre -Training Fine -Tuning

    Goal: Learn general image inpainting Goal: Learn intensity image appearance Self -supervised learning on RGB image datasets Real LiDAR data Accumulated scans as ground truth 7
  5. Method [3/4] Pre -Training: Image Inpainting with Readily Available Data

    Dataset and Online Masking Pre -Training Results Dataset: ~400,000 RGB images for VPR GSV-Cities [Ali -bey+, Neurocomputing 2022] Online Masking Examples Random Line Block 8
  6. Method [4/4] Fine -Tuning: Adapting to Sensor Data Collection -Specific

    Intensity Appearance Fine -Tuning Results 200 dense intensity images collected (0.05 % of the pre -training dataset) Livox AVIA Indoor Environments (Ouster OS1 -64) Livox MID360 Ouster OS1 -64 Outdoor Environments (Livox AVIA) 9
  7. Experiments and Evaluation [4/5] Experimental Results Higher relative pose accuracy

    than baselines without camera -LiDAR correspondence supervision. 10
  8. Conclusion Summary • Image -to -point cloud registration • Fine

    -tuning with limited LiDAR data via Rectified Flow -based LiDAR intensity upsampling ; no image -point cloud pairs or ground • Higher pose estimation accuracy than baselines -truth poses. in our evaluation. Future Work • More stable intensity generation in unstructured environments, such as trees and grass. • Integrate more advanced feature matchers, such as MASt3R. 11