Trustworthy 3D Vision
Revealing what 3D models learn and preserving what should remain private.
Hello! I’m Jaemin Cho.
I am a master’s candidate at Stony Brook University, advised by François Rameau.
My research centers on Spatial AI and Trustworthy 3D Vision, exploring the intersection of Visual Localization, 3D Foundation Models (e.g., DUSt3R, VGGT), and Model Privacy. My current interests lie in scene coordinate regression for visual localization, specifically focusing on model confidentiality and omnidirectional representation.
In addition to my research, I like to spend my downtime editing videos and experimenting with visual effects, having produced several independent films as a member of Inha University’s filmmaking club.
[New] Open to research opportunities starting January 2027. Please feel free to connect!
News
- We presented “Seeing Through the Weights” as a Long Oral at Malmö Arena, with invited PPVLM (talk) and PPCV (poster) presentations during ECCV 2026. 🇸🇪
- Our work, “Seeing Through the Weights,” was featured in Computer Vision News’ “ECCV Daily Saturday” issue at ECCV 2026. 📰
- Our paper, “Seeing Through the Weights: Privacy Leakage in Scene Coordinate Regression,” was accepted to ECCV 2026! ⭐
- Received the SUNY Korea Presidential Award with Oleksii for our research on privacy attacks in Scene Coordinate Regression.
- Started my M.S. in Computer Science at Stony Brook University, advised by François Rameau.
Highlights
Visual Localization & Mapping
Building reliable camera localization across diverse camera models and viewpoints.
Publications
Under Review
Ray-SCR: Generalizing Scene Coordinate Regression Across Camera Models
TL;DR: Ray-conditioned SCR for visual localization on cross-device (360Loc), omnidirectional (OmniScenes), and perspective (Aachen) datasets, achieving state-of-the-art results.
ECCV 2026
Long Oral
Seeing Through the Weights: Privacy Leakage in Scene Coordinate Regression
TL;DR: Recovered 3D scene and 2D images from Scene Coordinate Regression models for Visual Localization with domain agnostic images, challenging the SCR models are privacy-preserving.
