IbPRIA 2023 image
IbPRIA 2025: 12th Iberian Conference on Pattern Recognition and Image Analysis
Coimbra, Portugal. June 30 - July 3, 2025
Plenary Talks

Invited Speakers

Speaker Affiliation Photo Public page
António M. Lopez Autonomous University of Barcelona, Spain António M. López Google scholar link
Christoph Busch NTNU, Norway and HDA, Germany Christoph Busch Google scholar link
João Filipe Henriques University of Oxford, UK João Filipe Henriques Google scholar link
Luísa Proença Deputy National Director of Investigation Police, Portugal Luísa Proença LinkedIn link

Talks
[July 01 - 10h00] Challenges for Automated Face Recognition System
Prof. Christoph Busch
Abstract:
The talk will address challenges of face recognition systems. When dealing with operational systems, the quality of captured face images is relevant as it will impact the recognition accuracy. Thus, it is required to measure the utility of a face sample with a quality score but also with complementary measures that can provide actionable feedback. A serious challenge for face recognition systems is the vulnerability to presentation attacks for instance with silicon masks. For reliable recognition in non-supervised environments robust presentation attack detection is required. Further enrolment attacks that morph the face images of two subjects raised concerns. Such attacks merge the content of two parent images into one. This is problematic, as many countries still allow in the passport application analogue images, i.e., a printed photo. Last not least biometric templates must be protected. Acceptability of biometric systems requires fairness of biometric algorithms and artificial neural networks that are used. It is important to determine if face recognition systems are/are not biased towards a specific demographic group.

[July 02 - 10h00] Inner Thoughts: Interpreting Deep Networks with Causality and Tuning Contributions
Prof. João Filipe Henriques
Abstract:
Despite their enormous practical success, large deep neural networks are often treated like black boxes, since their high-dimensional internal representations do not lend themselves easily to direct interpretation. In this talk I will discuss two complementary views for understanding these internals.

The first one, focused on visual recognition and inspired by causal learning, analyses how CNNs can be constrained to guarantee that their internal representations will represent real physical variables, such as the positions of objects depicted in images. This makes them directly interpretable. In a somewhat surprising result, we exactly predict how the error of an unsupervised object detector changes with different architectural decisions.

The second one, focused on Large Language Models, inspects how pretraining and fine-tuning have very different influences on LLM responses to individual prompts. This allows us to decide whether a prompt is likely to elicit safe or unsafe responses, lets us steer model behaviour and attitudes by directly changing its internal representation, and allows us to test a hypothesis about how "jailbreaks" disable safety measures. Interpreting representations in terms of "fine-tuned" versus "pre-trained" then turns out to be broadly useful.

[July 02 - 14h30] Vision-based Autonomous Driving by Imitation Learning
Prof. António M. López
Abstract:
Developing autonomous vehicles requires training and testing AI drivers with supervised data gathered from a wide variety of driving scenarios. We could say that data is the driver in autonomous driving. This talk highlights the work carried out at CVC/UAB to reduce the need for manual data labeling, focusing on the use of sensorimotor models trained through imitation learning. Antonio’s team brings nearly seven years of experience to this field, ranging from simulation with CARLA to deploying real-world vehicles in the Catalan Pyrenees and on the UAB campus. Their research also includes comparative studies of human attention and AI driver attention. In this presentation, we will review the team's research journey on this topic from its beginnings to the present, discussing current achievements and open questions.

[July 03 - 10h30] Building Innovative AI-Driven Capabilities for Polícia Judiciária (Law Enforcement): from Research & Development to Regulatory Compliance
Dr. Luísa Proença and Dr. Filipe rodrigues
Abstract:
The Portuguese Criminal Police (Polícia Judiciária, PJ) is a Law Enforcement Agency primarily focused on the investigation of serious and organized forms of crime, including terrorism. The rapid pace of technological evolution in recent years has significantly transformed the modus operandi of criminal organizations, requiring PJ to constantly adapt and strengthen its operational capabilities to effectively respond to these emerging threats. In this context, PJ has been investing in Research & Innovation (R&I) initiatives aimed at creating new operational capabilities through advanced technologies, including the application of Artificial Intelligence (AI) in various investigative domains. This talk will present PJ's strategic approach to technological innovation, highlighting real-world examples of R&I projects and the development of AI-driven tools for operational use. It will also address the legal and regulatory challenges associated with deploying AI in law enforcement, particularly in light of the European Union's AI Act.