• Optimized a Structure-from-Motion pipeline resulting in a reduction of mapping times from 655min to 102min, a 6-fold speed up.
• Created benchmark tool that creates ground-truth maps and quantify the quality of mapping and localization with high accuracy.
• Wrote new production visual re-localization pipeline with PyTorch resulting in 3x faster, 10x higher throughput, using 100x less memory than previous baseline and saving more than 95% in cloud compute costs.
• Designed and implemented an end-to-end mapping pipeline that process raw data into production-ready SfM maps using Ray workflows on Azure Kubernetes Service (AKS).
• Trained a global feature model (NetVLAD) for image retrieval in specific data domains, resulting in 10% improvement in localization performance.
• Designed and implemented a system architecture using REDIS, Helm and K8s to allow scaling the re-localization service to a very large number of simultaneous requests across thousands of maps, whilst maintaining QoS.

• Ran several state-of-the-art SLAM methods on internal datasets and devised an evaluation protocol to compare them.
• Designed and developed a localization pipeline to align scans to a large scale reference reconstruction in challenging environments.
• Created a renderer that process >100bi points in less than 20s.
• Helped adding Lidar-based terms to COLMAP, reducing re-projection median errors from 60px to 3px.
• Trained and evaluated different NERF variants using aerial views.
• Performed Camera-IMU calibration which helped to identify time-sync issues.
• Created multiple Argo/K8s workflows to process data at scale using cloud infrastructure (GCP).

Used computer vision and machine learning techniques for augmented reality applications;

• Created a new benchmark and dataset for visual re-localization with a team of skilled researchers;
• Trained and evaluated several families of visual localisation methods on this benchmark;
• Used cloud infrastructure to train and evaluate deep learning models;
• Presented research outcomes to a cross-disciplinary audience;
• Paper published at ECCV 2022

• Prepared and delivered tutorials on PyTorch and fundamentals of machine learning;
• Prepared module assessments and marked students;
• Created a Dockerized Jupyterhub platform for students to run their code remotely.

The L3 Pilot European Consortium studies the impact of vehicle automation with key European OEMs. My main responsibilities were:

• Lead the creation of a responsive online platform to support data visualisation of piloting data across the project.
• Attend international meetings with key partners in the automotive sector to meet project requirements and report results.