Bachelor’s Degree

Vision in Practice: from the pixel to the first visual system

Target students and objectives

The pathway is intended for third-year bachelor’s students (maximum 12). Its objective is to consolidate the foundations of computer vision, images and video, through autonomous study of the laboratory material and direct hands-on practice in Python and OpenCV, and to offer, on a real project, a complete experience of the methodology of research: problem definition, choice of methods, quantitative evaluation, report, presentation. The pathway prepares the student for the transition to the master’s degree or professional life, leaving the student with documented work that can be incorporated into the thesis.

Structure of the pathway

Part A, programmed by the Department, comprises two self-study modules. A1 (30 hours), Image and video processing: from acquisition to color, histograms, filters, edges, blobs, fitting and alignment, up to recognition, segmentation, optical flow, tracking, neural networks, stereo, face recognition, scene understanding, pose and actions; OpenCV exercises and two intermediate verifications (low- and high-level) with quantitative evaluation. A2 (25 hours), System fundamentals and deep learning for vision: computer architecture and operating systems, databases and SQL, networks, cloud and IoT, with a mini case of a data acquisition and storage pipeline; convolutional networks, transfer learning and evaluation metrics, with the reproduction of a classification or segmentation on a public dataset. The workload of Part A is 30 + 25 = 55 hours.

Part B comprises two projects agreed upon with the student and the supervisor (a researcher in the laboratory who provides about one hour per week of support). P1 (30 hours): a comparative study on the robustness of classic features, SIFT, HOG, color, with respect to learned features, under noise and perturbations, on a sub-topic chosen by the student (natural images, video, or a medical dataset). P2 (30 hours): a choice from a menu of five options: face anonymization and privacy-preserving vision on public video; scene understanding or visual saliency on a chosen dataset; self-supervised visual anomaly detection on industrial inspection; data audit of bias and long-tail distribution with an active learning strategy; an immersive experience in augmented or virtual reality on a defined task (education, cultural heritage), with natural interaction and usability analysis. A further 10 hours for planning, reviews, report, and presentation. The workload of Part B is 30 + 30 + 10 = 70 hours.

The total workload is 55 + 70 = 125 hours per year, within the provided range. Successful completion is attested with a certificate from the Department; no credits are awarded.

Admission and assessment

Admission: third year, passed courses in computer vision, programming, and mathematics, average grade of at least 27 in the final year, letter of motivation, fifteen-minute interview. The commission (two professors from the Department and one researcher from the laboratory) examines the report, the code, and the presentation, and issues a “passed” or “not passed” outcome, which is recorded in the student’s file and is usable for thesis or scholarship selection.

Summary table of activities, bachelor’s pathway

ActivityTypeHours
A1, Image and video processing (2 intermediate verifications)Self-study30
A2, System fundamentals and deep learning for visionSelf-study25
P1, From classic features to learned featuresProject30
P2, Project from the laboratory menu (5 options)Project30
Planning, reviews, report and presentationProject10
Annual total 125

Master’s Degree

Trustworthy and efficient AI for vision and XR

Target students and objectives

The pathway is intended for first- and second-year master’s students (maximum 8). It deepens the laboratory’s frontier lines in efficiency of models and systems, trustworthy artificial intelligence, and multimodal vision, and leads students to complete projects with an element of novelty beyond mere reproducibility, with a view to a PhD program, a research master’s thesis, or industry R&D.

Structure of the pathway

Part A, programmed by the Department, comprises two self-study modules. B1 (30 hours), Architectures and efficiency: neural architecture search, distillation, pruning and quantization (FP16/INT8), adaptive precision, neuroscience-inspired architectures; on the systems side, custom kernels, vectorization, data layout for cache, graph-level and loop-level optimizations, support for accelerators; milestone: reduced-precision benchmark of a small visual model (accuracy, latency, memory). B2 (35 hours), Trustworthy and multimodal AI: machine unlearning (concept localization, causal routing, parameter-efficient suppression), GDPR and right to be forgotten, robustness to probing, extraction and membership inference, utility–forgetting trade-off, explainability and out-of-distribution evaluation; multimodal LLMs and visual knowledge graphs (KGs from images, image QA, graph consistency for hallucination detection, cross-modal retrieval); advanced video understanding (first and third person, cross-view alignment, cross-view learning); milestone: reproduction of a concept localization on a small transformer and an image–KG–QA pipeline on a reduced dataset. The workload of Part A is 30 + 35 = 65 hours.

Part B comprises two projects agreed upon with the student and the supervisor (about 1.5 hours per week, with access to the laboratory’s workstations and GPUs). P1 (35 hours, core of the pathway): making a visual model efficient, distillation and pruning, quantization, deployment on a GPU or an edge device, with a full benchmark of accuracy, latency, memory, and power consumption; optional extension: custom kernel. P2 (40 hours): a choice from four options, medical bioinformatics (EEG, ECG, fMRI or histopathology, with domain adaptation, multi-teacher uncertainty, Grad-CAM/SHAP for clinical validation); anomaly detection, visual (generative and self-supervised models, industrial inspection, domain shift, interpretability) or on signals (RF spectrum, time series, real time); advanced video (alignment between first and third person and cross-view on public datasets); immersive interaction (XR): a study on augmented or virtual reality, detection and mitigation of cybersickness, natural interaction, behavioural or cognitive analysis, applications to cultural heritage or medicine, with experimentation and usability analysis. A further 10 hours for planning, reviews, reports and presentations. The workload of Part B is 35 + 40 + 10 = 85 hours.

The total workload is 65 + 85 = 150 hours per year, within the provided range. Successful completion is attested with a certificate from the Department; no credits are awarded.

Admission and assessment

For each project, a report with a “novelty and limitations” section, code in a repository, and a fifteen-minute presentation; the commission examines the portfolio. Admission: grade of at least 27 in at least one course in machine learning, computer vision, or advanced programming (for newly enrolled students: bachelor’s thesis of at least 28), letter of motivation, interview. The commission (researchers from the laboratory) assesses scientific quality, the rigor of the evaluations, autonomy, reports, and presentations, and issues a “passed” or “not passed” outcome to be taken into account for thesis or PhD selection.

Summary table of activities, master’s pathway

ActivityTypeHours
B1, Architectures and efficiency (HPC, reduced precision)Self-study30
B2, Trustworthy and multimodal AI (unlearning, KG+LLM, video)Self-study35
P1, Human-Computer Interaction and XRSelf-study, Project35
P2, Domain project (medical / anomaly detection / video / XR)Project40
Planning, reviews, reports and presentationsProject10
Annual total 150