Manuel Marques

Recent papers

Hierarchical Spatio-Temporal Transformer for Coherent Emergency Department Forecasting

Filipa Lino, Bárbara Tavares, Carlos Santiago, Cláudia Soares, Manuel Marques

ECML PKDD Workshop, 2026

HierSTT is a hierarchical, Transformer-based deep learning framework designed to forecast emergency department (ED) patient demand consistently across three interconnected levels: individual hospitals, regional health administrations, and the national health system.

Spectral Guidance for Flexible and Efficient Control of Diffusion Models

Gabriel Moreira, Manuel Marques, João Paulo Costeira, Chenyan Xiong

ICML, 2026

github

Spectral Guidance is a training-free framework for controlling diffusion models by projecting guidance signals directly onto the intrinsic geometry of the generative process.

Native Hierarchical and Compositional Representations with Subspace Embeddings

Gabriel Moreira, Zita Marinho, Manuel Marques, João Paulo Costeira, Chenyan Xiong

ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2026

github

The paper proposes a paradigm shift in representation learning: representing concepts as linear subspaces rather than traditional vector points. This framework natively captures hierarchical inclusion and formal logical operations without relying on explicit supervision.

Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction

Luis Maria Perdigao, Miguel Costa, Carlos Santiago, Manuel Marques

IEEE Intelligent Transportation System Conference, 2026

githubdataset

We propose a gaze-guided deep learning framework that predicts perceived cycling safety by comparing street-view images and learning which environments people consider safer. By incorporating eye-tracking data into a Vision Transformer, the model aligns its attention with human visual fixations, producing safety predictions that are comparable to state-of-the-art methods while improving interpretability and providing insights for urban planning.

Face inpainting with Identity Preserving Latent Diffusion Models

Joao Santos, Carlos Santiago, Manuel Marques

arXiv (Cornell University), 2026

github

This work uses latent diffusion models to reconstruct occluded facial regions while preserving individual identity and realistic facial features. By incorporating identity-preserving constraints into the inpainting process, it aims to generate visually coherent and natural-looking face reconstructions without altering the person’s distinctive appearance.

Selected papers/works

2D Image head pose estimation via latent space regression under occlusion settings

José Celestino, Manuel Marques, Jacinto C. Nascimento, João Paulo Costeira

Pattern Recognition, 2022

github

We estimate a person’s head orientation from 2D images when face is partially occluded. This paper uses latent-space representation and regression to learn robust facial features and predict head pose more accurately despite occlusions, improving reliability in challenging real-world conditions.

3D Human Pose Estimation with Occlusions: Introducing BlendMimic3D Dataset and GCN Refinement

Filipa Lino, Carlos Santiago, Manuel Marques

IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024

githubproject pagedataset

The paper addresses the challenge of estimating 3D human body poses from visual data when body parts are partially or fully occluded. It introduces the BlendMimic3D dataset to support research under occlusion conditions and uses Graph Convolutional Network refinement to improve the consistency and accuracy of predicted 3D skeletal poses.

VICAN: Very Efficient Calibration Algorithm for Large Camera Networks

Gabriel Moreira, Manuel Marques, João Paulo Costeira, Alexander G. Hauptmann

IEEE International Conference on Robotics and Automation, 2024

githubproject page

A primal–dual bipartite pose-graph solver that jointly calibrates a marker object and estimates the extrinsic poses of every camera in the network from marker observations. VICAN calibrates a network with thousands of cameras in a few seconds.

Which Cycling Environment Appears Safer? Learning Cycling Safety Perceptions From Pairwise Image Comparisons

Miguel Costa, Manuel Marques, Carlos Lima Azevedo, Felix Wilhelm Siebert, Filipe Moura

IEEE Transactions on Intelligent Transportation Systems, 2025

github

The paper presents a deep learning approach to assess how safe different cycling environments appear to people by asking participants to compare pairs of real-world street images and select the safer one. It introduces a Siamese convolutional neural network that learns human safety perceptions, accounts for cases where both environments are considered equally safe, and generates scalable safety scores to help urban planners evaluate cycling infrastructure and prioritize improvements.

Hyperbolic vs Euclidean Embeddings in Few-Shot Learning: Two Sides of the Same Coin

Gabriel Moreira, Manuel Marques, João Paulo Costeira, Alexander G. Hauptmann

IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024

codegithub

This work compares hyperbolic and Euclidean embeddings for few-shot image classification, investigating how their geometric properties affect learning from limited examples. It shows that constraining embeddings to a fixed radius and using Euclidean distance can outperform hyperbolic approaches, particularly in high-dimensional settings.

Motion Analysis for People with Cerebral Palsy: A Vision Based Approach

Mário Macedo, Alexandre Candeias, Manuel Marques

International Conference on Rehabilitation Robotics, 2019

This work explores a computer vision-based approach to analyze the movements of people with cerebral palsy by extracting body poses and movement patterns from video recordings. It aims to provide a more accessible and objective alternative to traditional motion analysis methods, supporting the assessment of motor function and the monitoring of rehabilitation progress.

Vision Augmented Robot Feeding

Alexandre Candeias, Travers Rhodes, Manuel Marques, João Paulo Costeira, Manuela Veloso

ECCV Workshop, 2019

videoproject page

Exploring computer vision methodologies, we enable robots to assist people with feeding by identifying food, locating the user’s mouth, and guiding utensils toward the target. By integrating visual perception with robotic control, the system aims to provide safe, accurate, and adaptive feeding assistance, particularly for individuals with limited mobility or physical disabilities.

Recovery of Subspace Structure from High-Rank Data with Missing Entries

João Carvalho, Manuel Marques, João Paulo Costeira

International Conference on Image Processing, 2019

This work presents a method for reconstructing incomplete high-dimensional data by jointly estimating missing values and exploiting the underlying low-dimensional subspace structure. It improves data reconstruction and clustering accuracy, even when large portions of the data are missing, with applications including human pose recovery and image inpainting.