Understanding 3D structures from sketches

Sketches are a unique way to communicate: drawing a simple sketch does not require any training, sketches convey information that is hard to describe with words, they are powerful enough to represent almost any concept, and nowadays, it is possible to draw directly from mobile devices. Motivated from the unique characteristics of sketches and fascinated by the human ability to imagine 3D objects from drawings, this thesis focuses on automatically associating geometric information to sketches. The main research directions of the thesis can be summarized as obtaining geometric information from freehand scene sketches to improve 2D sketch-based tasks and investigating Vision-Language models to overcome 3D sketch-based tasks limitations. The first part of the thesis concerns geometric information prediction from scene sketches improving scene sketch to image generation and unlocking new creativity effects. The thesis proceeds showing a study conducted on the Vision-Language models embedding space considering sketches, line renderings and RGB renderings of 3D shape to overcome the use of supervised datasets for 3D sketch-based tasks, that are limited and hard to acquire. Following the obtained observations and results, Vision-Language models are applied to Sketch Based Shape Retrieval without the need of training on supervised datasets. We then analyze the use of Vision-Language models for sketch based 3D reconstruction in an unsupervised manner. In the final chapter we report the results obtained in an additional project carried during the PhD, which has lead to the development of a framework to learn an embedding space of neural networks that can be navigated to get ready-to-use models with desired characteristics.

Supervised and weakly supervised counting-by-segmentation: the fluorescent microscopy use case

This thesis focuses on automating the time-consuming task of manually counting activated neurons in fluorescent microscopy images, which is used to study the mechanisms underlying torpor. The traditional method of manual annotation can introduce bias and delay the outcome of experiments, so the author investigates a deep-learning-based procedure to automatize this task. The author explores two of the main convolutional-neural-network (CNNs) state-of-the-art architectures: UNet and ResUnet family model, and uses a counting-by-segmentation strategy to provide a justification of the objects considered during the counting process. The author also explores a weakly-supervised learning strategy that exploits only dot annotations. The author quantifies the advantages in terms of data reduction and counting performance boost obtainable with a transfer-learning approach and, specifically, a fine-tuning procedure. The author released the dataset used for the supervised use case and all the pre-training models, and designed a web application to share both the counting process pipeline developed in this work and the models pre-trained on the dataset analyzed in this work.

Directly training spiking neural networks for cyber-physical systems: from supervised to reinforcement learning

Spiking Neural Networks (SNNs) are bio-inspired Artificial Neural Networks (ANNs) utilizing discrete spiking signals, akin to neuron communication in the brain, making them ideal for real-time and energy-efficient Cyber-Physical Systems (CPSs). This thesis explores their potential in Structural Health Monitoring (SHM), leveraging low-cost MEMS accelerometers for early damage detection in motorway bridges. The study focuses on Long Short-Term SNNs (LSNNs), although their complex learning processes pose challenges. Comparing LSNNs with other ANN models and training algorithms for SHM, findings indicate LSNNs' effectiveness in damage identification, comparable to ANNs trained using traditional methods. Additionally, an optimized embedded LSNN implementation demonstrates a 54% reduction in execution time, but with longer pre-processing due to spike-based encoding. Furthermore, SNNs are applied in UAV obstacle avoidance, trained directly using a Reinforcement Learning (RL) algorithm with event-based input from a Dynamic Vision Sensor (DVS). Performance evaluation against Convolutional Neural Networks (CNNs) highlights SNNs' superior energy efficiency, showing a 6x decrease in energy consumption. The study also investigates embedded SNN implementations' latency and throughput in real-world deployments, emphasizing their potential for energy-efficient monitoring systems. This research contributes to advancing SHM and UAV obstacle avoidance through SNNs' efficient information processing and decision-making capabilities within CPS domains.