Internship - 3D Infant Motion Estimation from RGB Video

*Le descriptif de l’offre ci-dessous est en Anglais* **Type de contrat :** Convention de stage **Niveau de diplôme exigé :** Bac + 4 ou équivalent **Fonction :** Stagiaire de la recherche **A propos du centre ou de la direction fonctionnelle** ------------------------------------------------------- The Inria Grenoble research center groups together almost 600 people in 27 research teams and 8 research support departments. Staff is present on three campuses in Grenoble, in close collaboration with other research and higher education institutions (University Grenoble Alpes, CNRS, CEA, INRAE, …), but also with key economic players in the area. Inria Grenoble is active in the fields of high-performance computing, verification and embedded systems, modeling of the environment at multiple levels, and data science and artificial intelligence. The center is a top-level scientific institute with an extensive network of international collaborations in Europe and the rest of the world. **Contexte et atouts du poste** ------------------------------- The internship will take place within the Kinetix team at Inria of Grenoble Alpes University. The student will be supervised by Felix Küper, Nikolas Hesse, and Sergi Pujades. **Mission confiée** ------------------- **Context** Recovering 3D human pose and shape from images or videos is a core problem in 3D vision. Recent foundation models for human mesh recovery (HMR), such as SAM 3D Body [1], estimate full-body meshes from a single image with impressive robustness. Video-based methods add temporal consistency and handle camera motion. Large synthetic datasets such as BEDLAM2 [2] have been key to training them, because they provide perfect 3D ground truth. These models are trained almost exclusively on adults. Infants have different body proportions, poses, and appearance, and marker-based ground truth cannot be collected for them, so labeled data is scarce. Yet infant motion analysis has strong clinical value. The quality of spontaneous movement in the first months of life is an early indicator of neuromotor disorders such as cerebral palsy, and video-based markerless capture would make assessment possible at scale. Recent work [3] shows that fine-tuning SAM 3D Body on unlabeled infant video improves its predictions. However, it relies on pseudo-labels from another model, operates on single frames, and still reconstructs infants with an adult-shaped body model. **Internship** The goal of this internship is to improve the estimation of 3D infant motion from RGB video. The work will include: * running and benchmarking state-of-the-art HMR methods, image-based and video-based, on infant videos, and analyzing their failure modes; * creating a large-scale synthetic video dataset of infants with exact 3D ground truth, as an important milestone toward fine-tuning. To our knowledge, no such dataset exists; * running, fine-tuning, and improving video-based HMR methods, and evaluating the gains on rea