Towards CNN map representation and compression for camera relocalisation

TL;DR

This paper explores using CNNs for camera relocalisation and map compression, demonstrating that smaller CNN architectures can maintain accuracy while reducing map size, thus improving efficiency.

Contribution

It introduces a CNN-based map representation and compression method that preserves relocalisation performance with smaller network architectures.

Findings

Smaller CNNs can effectively represent maps for relocalisation.

Increasing training data improves accuracy without enlarging the CNN.

The approach achieves competitive results on public datasets.

Abstract

This paper presents a study on the use of Convolutional Neural Networks for camera relocalisation and its application to map compression. We follow state of the art visual relocalisation results and evaluate the response to different data inputs. We use a CNN map representation and introduce the notion of map compression under this paradigm by using smaller CNN architectures without sacrificing relocalisation performance. We evaluate this approach in a series of publicly available datasets over a number of CNN architectures with different sizes, both in complexity and number of layers. This formulation allows us to improve relocalisation accuracy by increasing the number of training trajectories while maintaining a constant-size CNN.