Projects
The research threads behind the papers, and the code that came out of them.
-
ARIQA-3DS — stereoscopic AR quality dataset
The first large stereoscopic AR image quality dataset: 1,200 viewports that fuse high-resolution stereoscopic omnidirectional captures of real scenes with augmented foregrounds under controlled transparency and degradation.
- Subjective study with 36 participants on a video see-through head-mounted display.
- Collects both quality ratings and simulator-sickness indicators.
- Released publicly as a benchmark for next-generation AR quality models.
-
TransformAR — lightweight AR quality metrics
A family of compact transformer metrics for augmented reality image quality — TransformAR, TransformAR-KD and TransformAR-KD+ — built so that perceptual accuracy survives aggressive model compression.
- Self-supervised ViT features distilled from reference to distorted images.
- Cross-attention decoders over shift representations to quantify distortion.
- Published at MMSP 2024, in Signal Processing: Image Communication, and in IEEE TCSVT.
-
Ranking-guided quality prediction
Humans rank image pairs far more consistently than they score single images. This project folds perceptual ordering feedback into the training objective of lightweight quality models, improving calibration and pairwise consistency.
-
Knee osteoarthritis severity grading
Transformer architectures that localise clinically salient biomarkers in knee radiographs to support Kellgren–Lawrence grading, with a localisation-aware refinement module that keeps local cues discriminative at every scale.
- Swin Transformer backbone with multi-head prediction (CBMI 2023).
- Localisation-aware token refinement for robust multi-dataset training (EUSIPCO 2024).
- Domain adaptation to mitigate shift between public radiographic datasets.
Code
github.com/AymenSe →Reference implementations for the papers, plus experiments that did not make it into one, are on GitHub. Anything released alongside a publication is linked from that paper’s page.