Spatial AI Lab Zurich: Advancing 3D Understanding and AI Interaction

Spatial AI Lab – Zurich

This document outlines the research conducted at the Spatial AI Lab in Zurich, a part of Microsoft Research. The lab focuses on advancements in artificial intelligence, particularly in areas related to spatial understanding and manipulation.

Overview

The Spatial AI Lab in Zurich is dedicated to pushing the boundaries of artificial intelligence, with a strong emphasis on how AI can understand, interact with, and generate the physical world. Their research spans various domains within AI, including computer vision, machine learning, and robotics.

Key Research Areas and Publications

The lab has a significant output of research papers, with a focus on computer vision and related fields. Notable areas of research include:

• 3D Reconstruction and Understanding:

  • No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images: This work explores generating 3D representations using Gaussian splats from minimal, unposed image data, suggesting a simpler approach to 3D reconstruction.
  • Robust Incremental Structure-from-Motion with Hybrid Features: Focuses on improving the robustness and efficiency of Structure-from-Motion (SfM) pipelines by integrating hybrid features.
  • 3D Neural Edge Reconstruction: Investigates neural methods for reconstructing 3D shapes by focusing on the critical edge information.
  • Sat2Scene: 3D Urban Scene Generation from Satellite Images with Diffusion: Utilizes diffusion models to generate 3D urban environments from satellite imagery.
  • F3Loc: Fusion and Filtering for Floorplan Localization: Presents a method for accurate localization using floor plans, combining sensor fusion and filtering techniques.
  • Multiway Point Cloud Mosaicking with Diffusion and Global Optimization: Explores the use of diffusion models and optimization for creating seamless point cloud mosaics.
  • EgoGen: An Egocentric Synthetic Data Generator: Develops a system for generating synthetic egocentric data, crucial for training AI models in robotics and autonomous systems.

• Neural Rendering and Representation:

  • GLACE: Global Local Accelerated Coordinate Encoding: Introduces a novel encoding method for neural representations, aiming for efficiency and quality in rendering.
  • Leveraging Neural Radiance Fields for Uncertainty-Aware Visual Localization: Applies Neural Radiance Fields (NeRFs) to visual localization tasks, incorporating uncertainty estimation.

• Computer Vision Techniques:

  • Structure-from-Motion from Pixel-wise Correspondences: Explores SfM using dense pixel-wise correspondences for improved accuracy.

People and Leadership

The lab is led by Marc Pollefeys, a Partner Director of Science. The research team includes numerous talented individuals contributing to these advancements. The website lists key researchers and their affiliations within Microsoft Research.

Collaboration and Engagement

The Spatial AI Lab actively engages with the research community through:

• Publications: Regularly publishing in top-tier computer vision and AI conferences like CVPR, ECCV, and ICLR.
• Projects: Showcasing ongoing research projects and their outcomes.
• Events: Participating in and hosting academic events and conferences.
• Connect & Learn: Sharing knowledge through podcasts, blogs, and forums.

Research Areas Filter

The website provides a filtering mechanism for research areas, allowing users to narrow down publications by categories such as:

• Computer vision (115 publications)
• Artificial intelligence (39 publications)
• Graphics and multimedia (11 publications)
• Systems and networking (2 publications)
• Human-computer interaction (2 publications)
• Algorithms (1 publication)

People Filter

Publications can also be filtered by author, highlighting key contributors like:

• Pablo Speciale (12 publications)
• Ondrej Miksik (6 publications)
• Mihai Dusmanu (6 publications)
• Remi Pautrat (6 publications)
• Silvano Galliani (5 publications)
• Christoph Vogel (5 publications)
• Rui Wang (4 publications)
• Jeffrey Delmerico (3 publications)
• Vibhav Vineet (2 publications)

Publication Types and Dates

Users can filter by publication type (e.g., Inproceedings, Article) and publication date ranges.

Contact and Social Media

The lab maintains a presence on social media platforms including X (formerly Twitter), Facebook, LinkedIn, and YouTube, facilitating communication and dissemination of research findings. Contact information and a physical address in Zurich are also provided.

Image

A prominent image associated with the lab is a background image, likely representing the visual or spatial nature of their work.

Overall Focus

The Spatial AI Lab in Zurich is a hub for cutting-edge research in spatial AI, contributing significantly to the fields of computer vision, 3D understanding, and AI-driven scene generation and manipulation. Their work is characterized by a strong publication record in leading conferences and a focus on practical applications of advanced AI techniques.