Giuseppe Facchi

3D Face Analysis Suite

The 3D Face Analysis Suite is an advanced platform for facial recognition, modeling, and evaluation that harnesses three-dimensional imaging to deliver precision beyond traditional 2D methods. It combines robust facial detection with accurate landmark tracking and detailed surface reconstruction, enabling reliable recognition across varying angles, lighting conditions, and expressions. By integrating depth information and texture analysis, the suite provides a highly realistic representation of facial features while also allowing for the interpretation of expressions and emotional states in real time. Designed with flexibility in mind, it supports seamless integration into diverse workflows, from biometric security and forensic investigation to medical diagnostics, surgical planning, making it a versatile solution for industries that require both accuracy and adaptability in facial analysis.

Syn-Detector: Automatic classification of facial syndromes with Graph Neural Networks

This project initially focuses on detecting Marfan syndrome using Graph Neural Networks (GNNs) applied to 3D facial data, exploiting the intricate geometry of facial structures to distinguish affected individuals. By leveraging GNNs, the approach captures spatial relationships that enhance diagnostic accuracy and reliability. In the future, this framework will expand to encompass a broader range of syndromic conditions, making it a versatile tool for multi-syndrome classification. This scalability allows for the potential integration of more complex, rare syndromes, providing valuable support for clinical decision-making in genetic and morphological research.

KA-GCN: Kernel-Attentive Graph Convolutional Network for 3D face analysis Graph Structure Learning

Graph Structure Learning (GSL) enhances Graph Neural Networks (GNNs) by refining graph structures for applications in unstructured domains like 3D face analysis, especially when data are limited. This paper introduces a kernel-attentive GCN (KA-GCN) that leverages an adaptive distance-based attention mechanism to learn adjacency matrices, demonstrating robust performance on both public and private datasets.

Spatial trajectory segmentation and analysis for determining stop points

University project of the Geospatial data management course focused on the implementation and comparison of three different spatial trajectories segmentation algorithms The segmentation is based on the well-known stop-move model and applied to data about the movements of persons in a museum. After the segmentation step, some analyses were performed on the result to find the most visited exhibits and the number of people standing together by the same exhibit using MobilityDB, a novel DBMS specialized in trajectory data management that supports data types representing values changing over time and specialized operations over them.

D415 GUI for Face Reconstruction

A GUI that implements a robust face reconstruction method from a single image, providing interactive features for users to visualize and manipulate 3D facial data efficiently.

Surgical Training video analysis

This project develops a tool for analyzing surgical training videos to enhance trainees' learning experiences. Using the YOLO (You Only Look Once) framework, it detects surgical instruments and key actions after the recording is complete, overlaying visual annotations to highlight performance. The tool incorporates template matching and SAM2 (Segment Anything Model 2) for effective object tracking, ensuring consistent monitoring throughout the procedure. Detected frames can be saved for focused review, and performance metrics like task completion times and instrument accuracy provide valuable feedback for progress tracking.

2S-SGCN: A Two-Stage Stratified Graph Convolutional Network Model for Facial Landmark Detection on 3D data

Facial Landmark Detection (FLD) has expanded into 3D, driven by its applications in fields like medical research, yet presents challenges in network architecture and precision. This paper introduces a 2-Stage Stratified Graph Convolutional Network (2S-SGCN) that uses heatmap regression and a new post-processing technique, achieving state-of-the-art results on public 3D datasets and supporting efficient deployment on resource-constrained devices.

BScObject detection and tracking in Surgical Training videos

The project focused on object detection and tracking in surgical training videos, using a hybrid approach that combined fine-tuned YOLO models, traditional computer vision techniques, and SAMURAI to achieve accurate detection and reliable tracking of surgical instruments for skill assessment and workflow analysis.

BScEvaluation of Prompt-derived MedSAM guided by Label Fusion

The study aims to evaluate the ability of MedSAM to replicate segmentation masks obtained from multiple (crowdsourced) annotations when guided by a geometric prompt derived from label fusion. The evaluation will be carried out without relying on ground truth, considering key factors such as inter-annotator variability (disagreement), the expertise level of the annotators, the estimated difficulty of each image, and the type of prompt provided to MedSAM.

BScEstimation of Fiducial Points on 3D Faces Using Multi-View Consensus-Based Techniques

The project aimed to develop a method for accurately estimating 478 fiducial points on 3D facial models using a multi-view approach. The process involved pre-processing the 3D mesh for proper alignment, generating orthographic views, and extracting 2D landmarks with MediaPipe, followed by a final estimation using a point density method. The resulting system demonstrated a strong balance between precision and computational efficiency, with identified areas for future improvement including enhanced support for various data formats and better handling of occluded points.

MScSegmentation of Choroid Plexus in MRI images: a new GT modality for the study of neurodegenerative diseases

The project aimed to implement SOTA methods in order to segment choroidal plexus from MRI images using T1 and FLAIR sequences. The segmentation was performed using U-Mamba, nn-UNet and Aschoplex architectures and the results were compared with the ground truth obtained from manual segmentation composed of T1 and FLAIR combinations.