Course Information

• When: Monday, June 19th, 2023

• Where: East 2

• Time: 8:30 AM - 12:15 PM (local time)

• Schedule

  • Introduction [5 min] [8:30 - 8:35]

  • Part I: Instance retrieval for coarse localization [45 min](Giorgos) [8:35 - 9:20] [slides]

  • Part II: Feature-based Visual Localization (Sudipta, Torsten, Zuzana) [9:20 - 11:05] 

    • Camera Pose Estimation [25 min] (Zuzana) [9:20 - 9:45]

  •  Coffee break [30 mins] [9:45 - 10:15]

    • Feature-based Visual Localization [30 min] (Torsten) [10:15 - 10:45]

    • Privacy-Preserving Localization [20 min] (Sudipta) [10:45 - 11:05]

  • Part III: Learning-based Visual Localization [40 min] (Eric) [11:05 - 11:45] [slides]

  • Part IV: Datasets & Benchmarking [30 min] (Marc) [11:45 - 12:15]

Material

Code pointers

• Retrieval and Coarse Localization

  • Classical

    • DenseVLAD

    • Aggregated Selective Match Kernel (ASMK) for retrieval

    • DisLocation

    • Place recognition with repetitive structures

  • Learning

    • NetVLAD

    • R-MAC and Integral Max pooling

    • CNN training for retrieval (GeM)

    • HOW deep local descriptors

    • Asymmetric metric learning (network)

    • Asymmetric metric learning (resolution)

    • Recall-at-k loss with very large batches

    • CosPlace 

• Visual Localization

  • Feature-based

    • kd-tree baseline, Vocabulary-based Prioritized Search (VPS), Active Search

    • 3D model compression

    • Geometric burstiness weighting for image retrieval

    • iOS demo for local pose tracking via SLAM

  • Learning-based

    • PoseNet

    • Relative Pose Regression (baselines of the Map-Free Relocalisation project)

    • Differentiable RANSAC

    • DSAC* (scene coordinate regression using RGB or RGB-D queries)

    • ESAC (scene coordinate regression on large scenes)

    • ACE (scene coordinate regression with fast mapping)

Datasets

• 7 Scenes (Visual Localization) - (alternative SfM ground truth poses available here)

• 12 Scenes (Visual Localization) - (alternative SfM ground truth poses available here) 

• Aachen (Visual Localization)

• Cambridge Landmarks (Visual Localization)

• Dubrovnik, Rome (Visual Localization)

• San Francisco SF-0, reference poses (Visual Localization)

• Aachen Day-Night and RobotCar Seasons datasets (Visual Localization) (dataset available upon request)

• TUM LSI Dataset (Visual Localization)

• Niantic Map-Free Dataset (Map-Free Visual Localisation)

• SfM120k training set for image retrieval

• Revisiting Oxford and Paris Benchmark

• Google Landmarks Dataset

• San Francisco eXtra Large (SF-XL)
  
  

Previous Related Tutorials

• ICCV 2021 Tutorial - Large-Scale Visual Localization

• ICCV 2019 Tutorial - Large-Scale Visual Place Recognition and Image-Based Localization

• ECCV 2018 Tutorial - Visual Localization: Feature-based vs. Learned Approaches

• CVPR 2017 Tutorial - Large-Scale Visual Place Recognition and Image-Based Localization

• CVPR 2015 Tutorial - Large-Scale Visual Place Recognition and Image-Based Localization

• CVPR 2014 Tutorial - Large-Scale Visual Place Recognition and Image-Based Localization