Zum Inhalt springen
Technische Universität München

News

Five CVPR 2026 Papers: Towards Agentic 3D Vision Systems

We are excited to announce five papers accepted to CVPR 2026, which will be presented at the main conference in Denver. Together, these works form a coherent research direction at the Professorship of Photogrammetry and Remote Sensing (PRS), TUM: from raw pixels → geometry → dynamics → agents. Our goal is to move towards general-purpose 3D vision systems that can see, understand, predict, and act in complex real-world environments.

🔹 Clean the Data

UnReflectAnything
Removes specular highlights from RGB images using synthetic supervision — enabling robust perception from imperfect real-world inputs.
👉 https://alberto-rota.github.io/UnReflectAnything/

🔹 Understand Objects (General & Zero-Shot)

ConceptPose
Training-free 6D pose estimation via language-driven concept reasoning — generalizes to unseen objects.
👉 https://stevenlk.xyz/conceptpose/

Object Pose Transformer (OPT-Pose)
Unified, category-agnostic model for unseen object pose estimation from RGB / RGB-D.
👉 https://colin-de.github.io/OPT-Pose/

🔹 Learn Dynamics from Observation

Node-RF
Continuous space-time modeling of scene dynamics using Neural ODEs — enables extrapolation beyond observed trajectories.
👉 https://around-star.github.io/node-rf/

🔹 Enable Agentic 3D Reasoning

VLM-Loc
Localizes in 3D maps from natural language using vision-language models and structured spatial reasoning.
👉 https://github.com/MCG-NKU/nku-3d-vision/tree/main/vlm-loc

💡 Together, these works move towards general-purpose 3D vision systems.
We look forward to presenting them at CVPR and discussing with the community.

◄ Zurück zu: News
To top