A half-day tutorial at the 18th Asian Conference on Computer Vision (ACCV 2026)
ACCV is a leading biennial international conference sponsored by the Asian Federation of Computer Vision.

Date: 14.12.2026
Time: AM
Room: 1009
Location: Osaka, Japan
Format: Half-day, in-person tutorial
Contact: alaa.othman@hsbi.de

Presented by: Dr. Alaa Tharwat
Bielefeld University of Applied Sciences and Arts, Bielefeld, Germany

[Registration / ACCV Passport Link] : https://accv2026.org/registration/

Why This Tutorial?

Modern computer vision models achieve impressive benchmark performance, yet they often fail silently when faced with out-of-distribution inputs, noisy data, or ambiguous scenarios. Knowing when a model’s prediction can be trusted is critical for deploying vision systems in safety-critical applications such as autonomous driving, medical diagnosis, and robotics.

Uncertainty Quantification (UQ) provides a principled framework for measuring and leveraging model confidence. This tutorial offers a comprehensive, hands-on introduction to UQ in computer vision. It covers theoretical foundations, classical Bayesian approaches, modern deep learning methods, and practical evaluation strategies. By the end, attendees will be equipped to build trustworthy vision systems that recognize their own limitations.

Overview

Computer vision models are increasingly deployed in high-stakes domains, but they lack the ability to say “I don’t know.” Uncertainty quantification addresses this gap by distinguishing between aleatoric uncertainty (inherent data noise) and epistemic uncertainty (model ignorance). This tutorial bridges theory and practice: from Bayesian inference and Gaussian Processes to Monte Carlo Dropout, Deep Ensembles, Evidential Deep Learning, and deterministic uncertainty methods.

The tutorial is designed for graduate students, researchers, and practitioners who want to apply UQ to real computer vision problems. It includes a hands-on coding session using PyTorch, case studies in medical imaging, autonomous driving, and remote sensing, and a discussion of calibration, out-of-distribution detection, and active learning.

What You Will Learn

By the end of this tutorial, participants will be able to:

• Explain the difference between aleatoric and epistemic uncertainty, and first-order vs. second-order probabilities.

• Understand classical Bayesian approaches: Bayesian inference, Gaussian Processes, Laplace approximation, and variational inference.

• Implement modern UQ methods for deep learning: Monte Carlo Dropout, Deep Ensembles, Evidential Deep Learning, and deterministic uncertainty methods (SnAP, DUQ, etc.).

• Evaluate uncertainty using calibration metrics such as Expected Calibration Error (ECE), Brier Score, and reliability diagrams.

• Apply uncertainty scores for out-of-distribution detection, active learning, and adversarial robustness.

• Visualize uncertainty maps for segmentation and object detection.

• Identify practical applications in medical imaging, autonomous driving, and remote sensing.

Schedule (tentative)

Half-day tutorial — 4 hours

Target Audience & Prerequisites

Target audience:

• Graduate students (MSc and PhD)

• Industry researchers

• Practitioners in computer vision, machine learning, and trustworthy AI

Prerequisites:

• Basic understanding of machine learning and deep learning

• Familiarity with training neural networks, classification, and regression

• Working knowledge of Python and a deep learning framework such as PyTorch

• Basic probability and statistics

Expected attendance: 50–100 participants

Organizer: Dr. Alaa Tharwat

Dr. Alaa Tharwat is a Postdoctoral Researcher and Research Group Leader in “Learning from Small Data” at the Bielefeld University of Applied Sciences and Arts, Germany. His research focuses on data-centric AI, with a specific emphasis on Active Learning, Uncertainty Quantification, and their applications in computer vision, remote sensing, and medical imaging. He has published extensively on these topics, including an award-winning survey paper on Active Learning (Tharwat & Schenck, 2023, Mathematics) that received the Best Paper Award. His work on uncertainty quantification explores how to measure and leverage model confidence for robust decision-making in safety-critical vision applications. He has organized multiple workshops and tutorials on data-efficient learning and uncertainty estimation at major international conferences, including ECML PKDD, IJCNN, and others.

Contact

For questions about the tutorial, please contact:

Dr. Alaa Tharwat
Bielefeld University of Applied Sciences and Arts, Bielefeld, Germany
Email: alaa.othman@hsbi.de