1 of 12

MeshVPR

Citywide Visual Place Recognition Using 3D Meshes

Gabriele Berton1, Lorenz Junglas2, Riccardo Zaccone1, Thomas Pollok3, Barbara Caputo1, Carlo Masone1

1Politecnico di Torino, 2Karlsruhe Institute of Technology, 3Fraunhofer IOSB

2 of 12

What is Visual Place Recognition

QUERY

52°30'N 13°22'E

OUTPUT

INPUT

DATABASE

INPUT

...

3 of 12

Applications of Visual Place Recognition

Disaster management

Journalism

Historians

Robotics

4 of 12

Deploying a VPR pipeline

Hardest part? Collecting a database

  • Expensive and time-consuming
  • Privacy concerns
  • Limited viewpoints
  • Sparse coverage

Solution? 3D meshes

5 of 12

VPR with 3D meshes

KNN

6 of 12

MeshVPR: training

MSE Loss

🔥

7 of 12

Datasets

8 of 12

Results

9 of 12

Results: mesh quality

10 of 12

Results: 3D meshes vs real images

11 of 12

12 of 12

MeshVPR

Citywide Visual Place Recognition Using 3D Meshes

Code + data + paper: https://meshvpr.github.io/

Gabriele Berton1, Lorenz Junglas2, Riccardo Zaccone1, Thomas Pollok3, Barbara Caputo1, Carlo Masone1

1Politecnico di Torino, 2Karlsruhe Institute of Technology, 3Fraunhofer IOSB

Tomorrow Demo 12:00-16:00 and poster 16:30-18:30