Tommaso Puccetti and Lorenzo Sarti (UNIFI-CINI) spent their one-month secondment at Universidade Estadual de Campinas, Brazil.

The activities carried out concern the evaluation of the robustness of a Visual Odometry system that entrusts the calculation of the position to two ML models: SuperGlue and SuperPoint.

To evaluate the system's robustness, a failure injection campaign was carried out on the vehicle's camera, simulating failures that can occur on its lens.

The picture below is from the seminar taken by Lorenzo and Tommaso regarding the results collected through the failure injection campaign, and proposals of the research roadmap for future collaborations.

The work carried out lays the foundations for a more extensive experimental campaign that could also include an additional system that combines the above-mentioned VO system with a component based on IMU sensors.

Furthermore, the long-term goal will be to test possible mitigations that can increase the robustness of the models against this type of failure. The first proposal is to evaluate the use of Self Supervised Learning techniques.

In this period, also Prof. Andrea Ceccarelli (UNIFI-CINI) was seconded for two weeks at UNICAMP, to guide and support Tommaso and Lorenzo in their activities.