Multi-Object Tracking
Unified and scalable multi-object tracker using hierarchical graph neural networks that process long clips across temporal scales.
Overview
We develop systems for understanding dynamic 3D scenes and temporal spatial relationships using multi-modal sensor data, such as cameras and LiDAR. Our research focuses on data-driven approaches that leverage temporal context to solve fundamental challenges in 4D perception: tracking objects across time, completing partial observations, and understanding scenes beyond fixed vocabularies. From graph-based multi-object tracking to feed-forward structure-from-motion, our methods enable robust spatial intelligence in real-world applications.
Points of Contact:
Laura Leal-Taixe,
Aljosa Osep,
Tim Meinhardt,
Guillem Braso,
Sergio Agostinho,
Qunjie Zhou
Key Research Areas
Detection, tracking, and segmentation • Structure-from-Motion • 4D Segmentation • Autolabeling
Featured Research Projects
Segment Anything in LiDAR
A unified framework for zero-shot LiDAR understanding that combines text-promptable 4D segmentation with feedforward object/scene completion.
Structure-from-Motion
End-to-end learnable framework for efficient large-scale Structure-from-Motion using novel latent global alignment with attention mechanisms.
ViPE: Video Pose Engine
Fast video engine that estimates camera intrinsics, camera motion, and dense near-metric depth from unconstrained videos. Robust across diverse scenarios and camera models.
Resources & Publications
- VGG-T3: Offline Feed-Forward 3D Reconstruction at Scale - CVPR 2026
- Towards Learning to Complete Anything in LiDAR - ICML 2025
- Zero-Shot 4D LiDAR Panoptic Segmentation - CVPR 2025
- Light3R-SfM: Towards Feed-forward Structure-from-Motion - CVPR 2025
- Better Call SAL: Towards Learning to Segment Anything in LiDAR - ECCV 2024
- Unifying Short and Long-Term Tracking with Graph Hierarchies - CVPR 2023