Traditional visual odometry (VO) methods are typically classified into feature-based and appearance-based approaches (also known as direct methods). In feature-based methods, the problem may be formulated as a windowed bundle adjustment (WBA) – or solved without it – minimizing the reprojection error. In direct methods, the objective is to define a cost function that estimates the camera pose by minimizing the photometric error across pixels.
Recently, deep visual odometry (DVO) approaches arise and they can be divided into modular and end-to-end methods. In modular approaches, DL is introduced into specific VO sub-tasks (e.g. feature extraction, feature matching, optical flow estimation, depth estimation or semantic understanding), where neural networks are highly specialized while the rest of the pipeline relies on classical geometry-based solutions. End-to-end DL solutions, where all tasks are learning-based, often suffer from poor generalization due to limitations in training datasets and typically incur high computational costs.
The PhD research should investigate how deep learning could support large-scale VO applications in real-world environments, considering indoor/outdoor complex scenarios, featuring textureless environments and illumination changes.
The expected developments and outcomes of the PhD include:
– Identifying the optimal balance between accuracy and processing time when choosing between a modular, component-wise integration and an end-to-end solution, while accounting for generalization limitations
– Developing methods for computational compression (e.g., neural network pruning, quantization, or distillation)
– Addressing the issue of poor generalization in deep learning by favouring local formulations (such as keypoint extraction and matching) that are inherently more transferable across environments
– Leveraging the retraining capabilities offered by the simulated environment of the ESA VAIPOSA project and studying how models trained in simulation can be effectively generalized to real-world datasets.
The ideal candidate should have an attitude towards problem-solving using programming environments, a strong background in AI and knowledge of robotics problems.