Niantic Spatial at ICCV 2025 — Honolulu, Hawaii 🌺
Niantic Spatial’s research team is proud to present significant advances at this year’s ICCV 2025 in Honolulu, part of our mission to build a living model of the world that people and machines can talk to.
Come meet us at the conference to discuss and find out more about our latest research and collaborations. A summary of our papers and projects is below 👇
1️⃣ ACE-G: Improving Generalization of Scene Coordinate Regression through Query Pre-Training
We introduce ACE-G, a two-fold approach to improve visual relocalization using scene coordinate regression. We use a scene-agnostic coordinate regressor and pair it with scene-specific map codes allowing us to pre-train the regressor specifically for generalization to query images.
Links: 📄 Paper | 💻 Project Page
2️⃣ PlaceIt3D: Language-Guided Object Placement in Real 3D Scenes
A large-scale benchmark and baseline (PlaceWizard) for translating natural language into grounded 3D object placements – advancing spatial reasoning for AR, robotics, and simulation.
Links: 📄 Paper | 💻 Project Page
3️⃣ Scene Coordinate Reconstruction Priors
Building on our previous work, ACE and ACE0, we model prior knowledge about the world and infuse it into ACE mapping. Our priors stabilize the reconstruction process, yielding better geometry for difficult scenes with positive effects on down-stream tasks.
Links: 📄 Paper | 💻 Project Page
We’re proud to contribute to the leadership of geospatial technology and productionized research at scale. If you’re at ICCV this week, come by and say hello.
See you in Honolulu! 🌴