R&D Engineer, Computer Vision and 3D Reconstruction

Context
This position is part of a project run by OnTheFly (onthefly3d.com), a startup venture spun out of Inria’s GraphDeco team, the researchers who invented 3D Gaussian Splatting, with support from Inria Startup Studio.
The project is building smartphone-based 3D capture tools for professional use, combining instant reconstruction with automated analysis of real-world sites over time.
Assignment
Develop methods to automatically compare 3D reconstructions of real-world captures against reference data, such as digital models or earlier captures of the same site. The work covers aligning heterogeneous 3D representations, as well as detecting and segmenting changes robustly under real-world capture conditions.
Main activities
Design and implement 3D alignment methods that combine visual or semantic correspondences with rendered geometric information. Develop geometry–material–lighting decomposition approaches for Gaussian Splatting representations. Build methods to detect and segment differences between captures, or between a reconstruction and its digital model, using images, depth, normals and vision models. Train and evaluate models on both real and synthetic data, then integrate the resulting methods into visualization tools.
Skills
Requirements
Engineering degree, Master's degree or PhD in Computer Science or related fields
Strong programming proficiency
Experience training and evaluating deep learning models with differentiable programming frameworks such as PyTorch
Solid grounding in computer vision, from fundamentals such as multiview geometry to modern vision models (e.g., vision transformers, vision-language models)
Interest in joining an early-stage startup
Preferred
Experience with shader or parallel programming (e.g., GLSL, CUDA)
Experience with 3D Gaussian Splatting, neural rendering or inverse rendering
Familiarity with BIM workflows