On 9 September 2026, researchers from ELLIS Unit Madrid presented new work at the Privacy, Fairness, Accountability and Transparency in Computer Vision (PFATCV) workshop, held in Malmö, Sweden, as part of ECCV 2026. The paper, “HuC-VideoMAE: Human-Centric Video Masked Autoencoding from Synthetic Data”, was selected for an oral presentation at the workshop.
PFATCV brings together researchers working on privacy, fairness, transparency, accountability and responsible AI in computer vision. As vision-based technologies become increasingly present in areas such as healthcare, robotics, surveillance and interactive systems, the workshop focuses on the technical and ethical challenges that arise from their deployment, including the protection of sensitive data, privacy risks and the responsible development of foundation models.
The paper, authored by Ricardo Pizarro, Roberto Valle, José M. Buenaposada, Luis M. Bergasa and Luis Baumela, investigates whether synthetic human-motion data can offer a more ethical alternative to the large collections of web-crawled videos commonly used to pretrain action-recognition models. The researchers propose HuC-VideoMAE, a human-centric masking strategy that uses body keypoints and person regions to help video transformers focus on the structure and dynamics of human motion. Their experiments show that this approach significantly improves pretraining with synthetic data, narrowing the performance gap with models trained on large-scale real-world video datasets without using a single real frame during pretraining.
José M. Buenaposada, Luis M. Bergasa and Luis Baumela are members of ELLIS Unit Madrid. The work contributes to an increasingly important research direction: developing computer vision systems that combine strong performance with greater attention to privacy, transparency and responsible AI. Its presentation at PFATCV offered a valuable opportunity to share these results with the ECCV community and discuss how synthetic data can support more ethical approaches to video understanding.