GeoUP Checkpoints

This repository contains the released checkpoints for GeoUP (Geometry-Grounded Unified 3D Perception for Autonomous Driving). GeoUP adapts the reconstruction-oriented latent representation of VGGT to calibrated, streaming, multi-camera driving scenes. A shared geometry-grounded representation supports metric depth estimation, 3D object detection, and semantic occupancy prediction.

The model code, configuration files, data preparation instructions, and evaluation scripts are available in the GeoUP GitHub repository.

Repository Structure

.
├── stage-1/
│   ├── low_resolution.pth
│   └── high_resolution.pth
└── stage-2/
    ├── nuscenes.pth
    └── multi-datasets.pth

Checkpoints

Checkpoint type	File	Intended use
GeoUP Pretrain Ckpts	stage-1/low_resolution.pth	Initialize GeoUP for low-resolution input recipes, including the nuScenes multi-task configuration.
GeoUP Pretrain Ckpts	stage-1/high_resolution.pth	Initialize GeoUP for high-resolution input recipes, including the joint multi-dataset configuration.
GeoUP Ckpts	stage-2/nuscenes.pth	Final nuScenes model for 3D detection, semantic occupancy prediction, metric depth estimation, and camera-pose evaluation.
GeoUP Ckpts	stage-2/multi-datasets.pth	Final jointly trained model for nuScenes, Argoverse 2, Waymo, KITTI, and DDAD.

GeoUP Pretrain Ckpts are initialization weights. Use them when starting training or fine-tuning with one of the maintained GeoUP configurations. GeoUP Ckpts are complete task models and should be used for evaluation or inference.

Citation

@article{xu2026geometry,
  title={Geometry-Grounded Unified 3D Perception for Autonomous Driving},
  author={Xu, Longfei and Wang, Xiaohui and Huang, Zehao and Li, Han and Yang, Ya and Wang, Naiyan and Liu, Si},
  journal={arXiv preprint arXiv:2608.13147},
  year={2026}
}

License

The checkpoints and associated code are released under the Apache License 2.0.