My current research focuses on ovarian cancer and aims to leverage all available patient data modalities : including CT and MRI imaging, histopathology whole-slide images, genomic profiles, and clinical variables, to improve survival prediction. Rather than treating each modality independently, the project investigates how modality-specific representations can be learned and combined to capture complementary and synergistic information.

This thesis aims to automatically select the most relevant 2D viewpoint for a given 3D object to facilitate identification and understanding. It quantifies viewpoint relevance by extracting essential object characteristics from available viewpoints. The research explores two main axes: evaluating the relevance of fixed viewpoints from textured images using geometric attributes and photographic recommendations, and determining the most representative viewpoint of an untextured 3D mesh by analyzing visible surface and intrinsic saliency based on viewing angle.

This paper introduces a new method for selecting the most relevant viewpoint of a 3D object, aiming to showcase the object effectively rather than focusing on aesthetics. The approach combines visibility and view-dependent saliency, incorporating factors like visible surface size and the saliency of visible vertices. A key contribution is an extensive evaluation protocol that includes a user study with over 200 participants and an analysis of image datasets to align with human preferences. The method outperforms existing techniques in selecting viewpoints that are most similar to human choices, even offering informative views where human biases might differ.
[Project Page]

This paper presents a method for automatically selecting the most relevant 2D viewpoint of a 3D object by quantifying visible "essential object characteristics" rather than aesthetics. This is a preliminary result of the main paper published at VISAPP 2024, mentioned above.

This paper introduces methods to quantify the relevance of 2D images in showcasing a 3D object, aiming to identify views that present essential characteristics regardless of texture. It proposes a deterministic relevance score based on geometric attributes and photographic recommendations, utilizing curvilinear saliency to extract features and filter irrelevant information. The paper also explores a learning-based approach using confidence scores from neural networks. Evaluation is performed using objective image rankings based on simulated degradations, demonstrating the deterministic method's efficiency and robustness, and providing insights into the behavior of learning-based methods.
[Project Page]

This poster, presenting results that earned the best contribution award at GTMG 2022, details a method for quantifying the visual relevance of a 2D image in showcasing a 3D object.

This paper proposes a method to quantify how well a 2D image showcases a 3D object, ranking images based on a "relevance score" derived from object dominance, size, and curvilinear saliency information. It also compares this approach to confidence scores from neural networks. Validated using systematically degraded images, the deterministic method proves effective and robust, outperforming most deep learning models. These works have been accepted at SCIA 2023.

This work is a preliminary result of the main paper published at SCIA 2023, which is mentioned above.

This paper proposes methods to quantify the essential 3D object information present in a 2D image, distinguishing between the "favorability" of an object's pose and the "relevance" of the 2D image itself. The core idea is to extract and match salient features from both 2D images and their corresponding 3D models using a curvilinear saliency detector, which ensures high repeatability between modalities. This allows for ranking poses and images based on the amount of essential information they contain about the object. The paper presents initial results that are encouraging, although it identifies areas for improvement such as handling occlusions and ensuring all salient points in depth maps are detected.