Research at the frontier of spatial intelligence.
Our teams in London, Oxford, across Europe and San Francisco publish at CVPR, ICCV, ECCV, NeurIPS and ICLR. The work ships into the models people and machines use every day.
Areas of focus
Reconstruction
Metric 3D from ordinary cameras. The line that began with Monodepth now reconstructs large outdoor scenes from street and satellite views.
Visual localization
Knowing exactly where a camera is from what it sees, from maps trained in minutes to models that work in places they have never seen.
Scene understanding
Structure and meaning in 3D, so people and agents can find planes, place objects and edit real scenes with language.
Spatial interfaces
Putting spatial intelligence to work for people, from guided scanning to navigation for blind and low-vision users.
Featured research
All publications
MVSAnywhere: metric depth for any scene, zero-shotFrom posed images, one general model predicts sharp, metric-accurate depth across indoor, outdoor and aerial scenes.
ACE-G: localization in places it has never seenPretrained across tens of thousands of real places, one model localizes in new scenes without retraining.
Cross-View Splatter: 3D from the street and the skyFuses satellite views with ground photos into Gaussian splats of large outdoor scenes, in one shared frame.Publications
47 publications
ECCV 2026AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak LabelsProject pageCVPR 2026Cross-View Splatter: Feed-Forward View Synthesis with Georeferenced ImagesProject pageICLR 2026A Scene is Worth a Thousand Features: Feed-Forward Camera Localization from a Collection of Image FeaturesProject pageCHI 2026NaviNote: Enabling In-situ Spatial Annotation Authoring to Support Exploration and Navigation for Blind and Low Vision PeopleProject pageICCV 2025ACE-G: Improving Generalization of Scene Coordinate Regression through Query Pre-TrainingProject pageICCV 2025PlaceIt3D: Language-Guided Object Placement in Real 3D ScenesProject pageICCV 2025Scene Coordinate Reconstruction PriorsProject pageUIST 2025ImaginateAR: AI‑Assisted In‑Situ Authoring in Augmented RealityProject pageUIST 2025AdjustAR: AI-Driven In-Situ Adjustment of Site-Specific Augmented Reality ContentProject pageCVPR 2025MVSAnywhere: Zero-Shot Multi-View StereoProject pageCVPR 2025Morpheus: Text-Driven 3D Gaussian Splat Shape and Color StylizationProject pageCHI 2025CoCreatAR: Enhancing Authoring of Outdoor Augmented Reality Experiences Through Asymmetric CollaborationProject pageECCV 2024HandDGP: Camera-Space Hand Mesh Prediction with Differentiable Global PositioningProject pageECCV 2024DoubleTake: Geometry Guided Depth EstimationProject pageECCV 2024Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a RelocalizerProject pageCHI 2024Don't Look Now: Audio/Haptic Guidance for 3D Scanning of LandmarksProject pageCVPR 2024Map-Relative Pose Regression for Visual Re-LocalizationProject pageCVPR 2024AirPlanes: Accurate Plane Estimation via 3D-Consistent EmbeddingsProject pageCVPR 2024Matching 2D Images in 3D: Metric Relative Pose from Metric CorrespondencesProject pageCVPR 2023Virtual Occlusions Through Implicit DepthProject pageCVPR 2023Removing Objects From Neural Radiance FieldsProject pageCVPR 2023Two-view Geometry Scoring Without CorrespondencesCodeCVPR 2023DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion ModelsCodeCVPR 2023Accelerated Coordinate Encoding: Learning to Relocalize in Minutes using RGB and PosesProject pageWACV 2023Heightfields for Efficient Scene Reconstruction for ARProject pageECCV 2022SimpleRecon: 3D Reconstruction Without 3D ConvolutionsProject pageECCV 2022Camera Pose Estimation and Localization with Active Audio SensingProject pageECCV 2022Map-free Visual Relocalization: Metric Pose Relative to a Single ImageProject pageCVPR 2022Learning Structured Gaussians to Approximate Deep EnsemblesProject page3DV 2021Visual Camera Re-Localization using Graph Neural Networks and Relative Pose SupervisionCodeICCV 2021On the Limits of Pseudo Ground Truth in Visual Camera Re-LocalisationCodeCVPR 2021Cuboids Revisited: Learning Robust 3D Shape Fitting to Single RGB ImagesCodeCVPR 2021Single Image Depth Prediction with Wavelet DecompositionCodeCVPR 2021The Temporal Opportunist: Self-Supervised Multi-Frame Monocular DepthCodeCVPR 2021Panoptic Segmentation ForecastingCodeICRA 2021Learning to Predict Repeatability of Interest PointsCodeECCV 2020Learning Stereo from Single ImagesProject pageNeurIPS 2020Physically Plausible Multi-Object Scene Synthesis Using Structured Latent SpacesCodeECCV 2020Predicting Visual Overlap of Images Through Interpretable Non-Metric Box EmbeddingsCodeECCV 2020Single-Image Depth Prediction Makes Feature Matching EasierProject pageCVPR 2020Footprints and Free Space from a Single Color ImageProject pageICCV 2019Self-Supervised Monocular Depth HintsCodeICCV 2019Digging Into Self-Supervised Monocular Depth EstimationCodeBMVC 2019An Efficient 3D CNN for Action/Object Segmentation in VideoProject pageSIGGRAPH 2019iMapper: Interaction-guided Scene Mapping from Monocular VideosCodeACCV 2018Unsupervised Intuitive Physics from Visual ObservationsProject pageCVPR 2017Unsupervised Monocular Depth Estimation with Left-Right ConsistencyCode
No publications match.
Work on the hardest problems in physical AI, with the team that mapped the world.Open roles