IPAB Workshop - 15/05/2025 Speaker: Bruce Zhang Title: One-shot Adaptive Mesh Generation for 3D Poisson Equations Abstract: We propose a fast, neural-based approach for mesh size field prediction that enables one-shot adaptive refinement for solving partial differential equations (PDEs). Unlike traditional adaptive mesh refinement (AMR), which requires repeated PDE solves, our method predicts the size field in a single step, significantly reducing computational cost while achieving comparable solution accuracy. We show that even simple neural architectures generalize well across varied boundary conditions and mesh geometries in small-scale settings. For larger and more diverse datasets, our approach scales effectively with increased model capacity, demonstrating strong potential for efficient mesh adaptation in large-scale scientific computing workflows. Speaker: Florent Le Moel Title: A Modular Software Platform for Multi-Camera 3D Tracking of Animal Behaviour Abstract: Accurate 3D reconstruction of animal poses is key for understanding many aspects of animal behaviour, yet it remains technically challenging. Although the geometry of triangulating 2D detections from multiple camera views is conceptually straightforward, practical issues such as camera synchronisation, calibration quality, and multi-animal identity association significantly hinder widespread use of 3D tracking in behavioural research. These challenges are even bigger in studies involving very small animals like insects, where even minor errors can result in substantial inaccuracies relative to their body sizes. We present a modular software platform that addresses these challenges through a robust pipeline for 3D pose estimation and interaction analysis from multi-camera video. The platform offers (1) acquisition tools supporting various camera types, synchronisation methods, and video formats; (2) precise, real-time camera calibration routine with automated spatial sampling and iterative refinement for intrinsic and extrinsic parameters; and (3) a multi-object tracking method for association of noisy, non-proofread 2D detections across time and views to enable accurate triangulation. Designed to integrate with existing 2D pose estimation and segmentation tools (e.g. DeepLabCut, SLEAP, SAM2), the system enables 3D reconstruction of multiple animal poses and rigid object geometry. We demonstrate the system’s effectiveness by recording free-moving ants manipulating objects, reconstructing their 3D body pose and interactions with millimetre-scale precision. May 15 2025 13.00 - 14.00 IPAB Workshop - 15/05/2025 Bruce Zhang & Florent Le Moel G.03
IPAB Workshop - 15/05/2025 Speaker: Bruce Zhang Title: One-shot Adaptive Mesh Generation for 3D Poisson Equations Abstract: We propose a fast, neural-based approach for mesh size field prediction that enables one-shot adaptive refinement for solving partial differential equations (PDEs). Unlike traditional adaptive mesh refinement (AMR), which requires repeated PDE solves, our method predicts the size field in a single step, significantly reducing computational cost while achieving comparable solution accuracy. We show that even simple neural architectures generalize well across varied boundary conditions and mesh geometries in small-scale settings. For larger and more diverse datasets, our approach scales effectively with increased model capacity, demonstrating strong potential for efficient mesh adaptation in large-scale scientific computing workflows. Speaker: Florent Le Moel Title: A Modular Software Platform for Multi-Camera 3D Tracking of Animal Behaviour Abstract: Accurate 3D reconstruction of animal poses is key for understanding many aspects of animal behaviour, yet it remains technically challenging. Although the geometry of triangulating 2D detections from multiple camera views is conceptually straightforward, practical issues such as camera synchronisation, calibration quality, and multi-animal identity association significantly hinder widespread use of 3D tracking in behavioural research. These challenges are even bigger in studies involving very small animals like insects, where even minor errors can result in substantial inaccuracies relative to their body sizes. We present a modular software platform that addresses these challenges through a robust pipeline for 3D pose estimation and interaction analysis from multi-camera video. The platform offers (1) acquisition tools supporting various camera types, synchronisation methods, and video formats; (2) precise, real-time camera calibration routine with automated spatial sampling and iterative refinement for intrinsic and extrinsic parameters; and (3) a multi-object tracking method for association of noisy, non-proofread 2D detections across time and views to enable accurate triangulation. Designed to integrate with existing 2D pose estimation and segmentation tools (e.g. DeepLabCut, SLEAP, SAM2), the system enables 3D reconstruction of multiple animal poses and rigid object geometry. We demonstrate the system’s effectiveness by recording free-moving ants manipulating objects, reconstructing their 3D body pose and interactions with millimetre-scale precision. May 15 2025 13.00 - 14.00 IPAB Workshop - 15/05/2025 Bruce Zhang & Florent Le Moel G.03
