Fully Convolutional Place Recognition Network

In the scope of the course “3D Vision” at ETH, I developed a novel deep network architecture (temporarly named “Fully-Convolutional Place Recognition Network”) to efficiently solve the problem of place recognition using only point-clouds as input.

This approach aims to solve the main issues coming from performing place recognition on images, such as the great variability induced by lighting and weather condition changes. However, there are many challenges in designing a solution able to extract efficiently meaningful features from point-clouds, and that can scale well at large scale.

The final solution showed promising performance, close the state of the art, but since more accurate evaluations are still in progress, not all details can be shared yet.