ThinkProbe: Beyond Accuracy, Structural Profiling of LLM Reasoning Traces
When we evaluate a reasoning model, we almost always look at one thing: was the final answer right? In ThinkProbe, my collaborators and I ask a different question: not whether...
Closing the Foveal Gap: Perceptually Grounded Scanpath Comparison with Disc IoU
Fixations are usually treated as exact points on an image, but human vision does not work that way. Each fixation gathers information from a foveal region around the gaze point....
What They Saw, Not Just Where They Looked: Semantic Scanpath Similarity via VLMs and NLP Metrics
Classic scanpath metrics tell us where two people looked and how well their gaze paths overlap in space. In this work, presented at ETRA 2026, we ask a different question:...
SPGen: Stochastic Scanpath Generation for Paintings using Unsupervised Domain Adaptation
SPGen generates diverse, human-like scanpaths for paintings, capturing not one canonical viewing path but the variety of ways people explore a work of art. We frame the problem stochastically so...
Morphology-Aware KOA Classification: Integrating Graph Priors with Vision Models
Grading knee osteoarthritis (KOA) severity from X-rays is a subtle task where the structure of the joint carries much of the diagnostic signal. In this work, we combine vision models...
A Graph-Driven Approach to Knee Osteoarthritis Severity Classification
Knee osteoarthritis severity classification benefits from reasoning about structure, not just appearance. In this work, presented at EUSIPCO 2025, we take a graph-driven approach to KOA grading, representing the structural...
Modeling Beyond MOS: Quality Assessment Models Must Integrate Context, Reasoning, and Multimodality
Perceptual quality assessment has long been organized around a single number: the Mean Opinion Score. In this perspective piece, we argue that this target is too narrow, and that the...
Shifts in Doctors' Eye Movements Between Real and AI-Generated Medical Images
As generative models produce increasingly convincing medical images, a natural question follows: do clinicians look at synthetic images the same way they look at real ones? In this ETRA 2025...
APEX: Agentic Portrait Editing
APEX is an experimental agentic pipeline for editing portraits from natural-language instructions. The idea is to let an AI agent sit between a user’s plain-language request and the image editing...
llm-tuning-lab: Recipes for Fine-Tuning Small Language Models
llm-tuning-lab is my personal lab for experimenting with the fine-tuning of small language models. It is meant for practitioners and researchers who want a place to collect working recipes, sanity-check...
Enlighten-AI: A RAG-Based Coaching Assistant
Enlighten-AI is a personal, experimental project where I built a retrieval-augmented generation (RAG) assistant with a supportive, coaching-style voice. It’s for anyone curious about how RAG can ground a language...
Shifting Focus: From Global Semantics to Local Prominent Features in Swin-Transformer for Knee Osteoarthritis Severity Assessment
Grading knee osteoarthritis (KOA) severity from radiographs is a diagnosis that hinges on small, localized cues (joint space narrowing, osteophytes, subtle bony changes), yet many transformer models are pulled toward...
Perceptual Evaluation of Masked AutoEncoder Emergent Properties Through Eye-Tracking-Based Policy
Masked autoencoders (MAEs) learn rich visual representations without labels, and they exhibit intriguing emergent properties along the way. But how do we judge whether what they learn aligns with human...
AVAtt: Art Visual Attention Dataset for Diverse Painting Styles
Where do people look when they view a painting, and how does that change across artistic styles? AVAtt, introduced at ETRA 2024, is a dataset of human visual attention over...
HeReFaNMi: Health-Related Fake News Mitigation
HeReFaNMi is a project aimed at detecting and mitigating health-related misinformation using natural language processing. Medical fake news is a particularly high-stakes domain (bad information about treatments, vaccines, or symptoms...
Quantization Effects on Neural Networks Perception
Quantization is one of the most practical tools we have for shrinking vision models and running them on constrained hardware, but it is almost always judged by a single yardstick:...
PhD Thesis Defended - A Gaze into the Art World
I’m thrilled to announce that I have successfully defended my PhD thesis!
Insights into Classifying and Mitigating LLMs' Hallucinations
Large language models are remarkably fluent, and that fluency is precisely what makes their hallucinations dangerous: confidently stated content that is simply not grounded in fact. In this work, we...
ICIP 2023 Point Cloud Visual Quality Assessment Grand Challenge
As 3D point clouds become central to immersive media and capture pipelines, we need dependable ways to measure their perceptual visual quality. I took part in the ICIP 2023 Point...
Automatic Diagnosis of Knee Osteoarthritis Severity Using Swin Transformer
Assessing knee osteoarthritis (KOA) severity from radiographs is a subtle, labor-intensive task, and consistent grading is hard even for experts. In this CBMI 2023 paper, we apply the Swin Transformer...
Detecting Colour Vision Deficiencies via Webcam-Based Eye-Tracking: A Case Study
Colour vision deficiencies are common, yet screening for them still often depends on dedicated tests or clinical settings. In this ETRA 2023 case study, we explore whether affordable webcam-based eye-tracking...
mlflow_pytorch_exp: Reproducible MLflow + PyTorch Boilerplate
mlflow_pytorch_exp is a boilerplate and scaffold generator for starting MLflow + PyTorch experiments. It is aimed at anyone who has felt the friction of setting up experiment tracking from scratch...
Deep-Based Quality Assessment of Medical Images Through Domain Adaptation
Assessing the quality of a medical image is a clinical necessity: artefacts, noise, and degradations can obscure findings and mislead diagnosis. In this work, presented at ICIP 2022, we tackle...
Paper Published - Domain Adaptive Deep Learning for Scanpath Prediction
Our latest paper “A Domain Adaptive Deep Learning Solution for Scanpath Prediction of Paintings” is now available on arXiv!
Paper at CVPR 2022 Workshop - Self-Supervised Scanpath Prediction for Paintings
Thrilled to announce that our paper on self-supervised scanpath prediction for painting images has been published at a CVPR 2022 Workshop!
SATSal: A Multi-Level Self-Attention Based Architecture for Visual Saliency Prediction
Predicting where people look in an image is a deceptively hard problem: attention is shaped by both fine local detail and broad scene structure. In SATSal, we introduce a saliency-prediction...
A Simple and Efficient Deep Scanpath Prediction
A scanpath is the sequence of fixations that traces how a viewer’s gaze moves across an image over time. Predicting it well usually invites heavy, elaborate architectures. In this work,...
SalyPath360: Saliency and Scanpath Prediction Framework for Omnidirectional Images
Omnidirectional, or 360°, images pose a distinct challenge for models of human attention. The viewer is placed at the center of a full sphere, free to look in any direction,...
Paper Published: Salypath - A Deep-Based Architecture for Visual Attention Prediction
Excited to share that our paper “Salypath: A Deep-Based Architecture For Visual Attention Prediction” has been published on arXiv!
Starting My PhD Journey
I’m excited to announce that I’ve started my PhD at the PRISME Laboratory, University of Orleans, France!