Student project: Tree Skeletonization in AI-Based 3D Reconstruction for Robotics

Student project: Tree Skeletonization in AI-Based 3D Reconstruction for Robotics *Important for non-EU students: You'll need to be registered at a Dutch university to meet immigration requirements. Extract hidden tree structure directly from AI-based 3D Gaussian Splatting by transforming raw gaussians into skeleton nodes for robotics tasks in orchards. What you will do Orchard operations such as pruning require detailed information about branches, junctions, and growth patterns, which can often be observed in images but may be lost or simplified in conventional 3D reconstructions. Tree skeletonization can provide a compact structural representation of these branches, helping to preserve and analyze the fine geometric details needed for pruning and other orchard operations. This project focuses on exploring methods to extract tree skeletons directly from 3D Gaussian Splatting representations. By using gaussian attributes such as position, scale, orientation, opacity, covariance, and anisotropy, the project aims to identify branch-related gaussians and organize them into a tree skeleton for plant analysis, phenotyping, and robotic applications. While methods that avoid conversion to point clouds are preferred, point cloud-based skeletonization can still serve as a useful baseline for comparison. Literature review: Review existing work on 3D Gaussian Splatting, plant/tree reconstruction, and tree skeletonization methods. Timeline with deliverables: Define a project plan with clear deliverables for design, implementation, integration, evaluation, and reporting. 3D Gaussian Splatting processing: Reconstruct or use an existing 3DGS tree model and apply preprocessing and postprocessing to improve the gaussian representation. Gaussian attribute analysis: Analyze gaussian attributes such as position, scale, rotation, opacity, color, covariance, anisotropy, local density, and neighborhood consistency to identify useful structural cues. Structural Gaussian selection: Develop a strategy to select or weight the Gaussians that are most relevant for recovering the tree skeleton. Skeleton extraction: Design and implement a method to create a graph from some specific gaussians and extract a tree skeleton. Evaluation and insights: Compare the proposed method with state-of-the-art tree skeletonization baseline and analyze the benefits and limitations of using gaussian attributes. What we do for you We have a challenging problem where you have a lot of freedom to come up with solutions. We have a diverse team of experts from the computer vision and AI fields to supervise and support you. You will join the Near-Sensor AI team of OnePlanet, which employs state of the art knowledge on computer vision, robotics, and AI for precision agriculture. You will be able to exchange views and knowledge with the OnePlanet and Imec community of experts and scientists, widening your professional network. We can help you to improve your coding skills up to industry standards. You have access to our cloud solutions to solve this problem allowing you to process large amount of data within reasonable time. Who you are Required skills: Proficiency in Python for data processing, algorithm development, and visualization. Basic knowledge of computer vision and 3D data processing. Familiarity with 3D representations such as point clouds, voxels, or gaussian-based representations. Basic knowledge of linear algebra, geometry, and multivariate statistics. Ability to work independently, explore open research problems, and report results clearly. Proficiency in using Git for version control. Interest in applying AI and computer vision to environmental, agricultural, or societal challenges. Preferred skills: Experience with 3D Gaussian Splatting. Knowledge of tree skeletonization methods. Experience with supervised learning. Basic understanding of Agile/Scrum methodologies. Interested Does this position sound like an interesting next step in your career at imec? Don’t hesitate to submit your application by clicking on ‘APPLY’. Should you have more questions about the job, you can contact jobs@imec.nl