From Privacy to Robustness in Computer Vision at BMVC 2024

12/02/2024

With the 35th British Machine Vision Conference in the books, time for a quick recap of our contributions to BMVC 2024 in the dynamic city of Glasgow, Scotland.

Our Contributions

At BMVC 2024, our lab contributed through two impactful presentations, each highlighting our team's ongoing efforts in computer vision and medical image analysis.

Francesco Di Salvo presented a poster at the main conference titled "Privacy-preserving datasets by capturing feature distributions with Conditional VAEs." This research addresses the growing concern of privacy in data sharing, particularly in the medical domain. By introducing a novel approach using Conditional Variational Autoencoders (CVAEs), Francesco’s work demonstrated how large, diverse, and privacy-preserving synthetic datasets can be generated. This method outperformed traditional anonymization techniques in both medical and natural image domains, showcasing its ability to enhance dataset diversity and model robustness while maintaining privacy.

Sebastian Doerrich delivered an oral presentation at the Robust Recognition in the Open World Workshop with his paper titled "Unsupervised Feature Orthogonalization for Learning Distortion-Invariant Representations." In this work, Sebastian introduced unORANIC+, a method that integrates unsupervised feature orthogonalization with Vision Transformers to improve robustness and generalizability. The technique separates anatomical and image-specific attributes to create robust latent representations, demonstrating excellent performance across diverse medical imaging tasks, datasets, and distortions. Sebastian’s work showed that unORANIC+ holds great potential for medical image analysis, especially in environments where large, tailored datasets are scarce.

See more details here.

To BMVC

The British Machine Vision Conference (BMVC) is an annual event organized by the British Machine Vision Association (BMVA). BMVC brings together researchers, practitioners, and industry leaders from around the world, spanning experties in computer vision, image processing, and pattern recognition. The 35th edition, held in Glasgow, Scotland, from November 25th to 28th, 2024. This year, BMVC received an impressive 1,020 submissions, with 264 papers accepted, underscoring the conference’s reputation for high-quality research and rigorous standards.

For Sebastian and Francesco, BMVC 2024 was a fantastic opportunity to present their work, exchange ideas, and build meaningful connections with fellow PhD students and experts from both industry and academia across the globe.