---
license: cc-by-nc-sa-4.0
datasets:
  - shdragron/RoboGeo
tags:
  - autonomous-driving
  - bev-perception
  - multi-camera
  - robustness
  - 3d-object-detection
  - bev-segmentation
---

# RoboGeo checkpoints

Trained checkpoints for **RoboGeo: Benchmarking Camera-Geometry Robustness in Multi-Camera BEV Perception**
(WACV 2027) · [Project page](https://robogeo2027.github.io/) · [Dataset](https://huggingface.co/datasets/shdragron/RoboGeo)

Each model is trained on one platform of RoboGeo. The Sedan models are evaluated under viewpoint
perturbation and cross-platform transfer; the SUV and Bus models are the models trained on the target
platforms of cross-platform transfer.

## 3D object detection

| model | platform | file | size | sha256 (first 16) |
|---|---|---|---|---|
| BEVFormer-tiny | Sedan | `detection/bevformer_tiny/bevformer_tiny_sedan.pth` | 134 MB | `e166f0aec95b1cff` |
| BEVFormer-tiny | SUV | `detection/bevformer_tiny/bevformer_tiny_suv.pth` | 134 MB | `68192ad4ce650624` |
| BEVFormer-tiny | Bus | `detection/bevformer_tiny/bevformer_tiny_bus.pth` | 134 MB | `2e5370a87ea68f81` |
| CAPE | Sedan | `detection/cape/cape_sedan.pth` | 245 MB | `664d2d6f53303c5f` |
| CAPE | SUV | `detection/cape/cape_suv.pth` | 245 MB | `e09f1e8893512ecf` |
| CAPE | Bus | `detection/cape/cape_bus.pth` | 245 MB | `9b2d48e700db18ae` |

## BEV segmentation

| model | platform | file | size | sha256 (first 16) |
|---|---|---|---|---|
| GaussianLSS | Sedan | `segmentation/gaussianlss/gaussianlss_sedan.pth` | 33.8 MB | `4ca7cf0901b99c34` |
| GaussianLSS | SUV | `segmentation/gaussianlss/gaussianlss_suv.pth` | 33.8 MB | `087171e00240fcaf` |
| GaussianLSS | Bus | `segmentation/gaussianlss/gaussianlss_bus.pth` | 33.8 MB | `149d3bd36c34cd17` |
| CVT | Sedan | `segmentation/cvt/cvt_sedan.pth` | 4.5 MB | `ded6a2c934a89edb` |
| CVT | SUV | `segmentation/cvt/cvt_suv.pth` | 4.5 MB | `946356200d2cdff1` |
| CVT | Bus | `segmentation/cvt/cvt_bus.pth` | 4.5 MB | `257bdae34378407e` |

Full checksums are in `SHA256SUMS`. Detection checkpoints hold `state_dict` and `meta.CLASSES`;
segmentation checkpoints hold `state_dict`. All load with `torch.load(path, weights_only=True)`.
Training and evaluation code: [shdragron/RoboGeo](https://github.com/shdragron/RoboGeo).

## License

Released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/), the license of the
RoboGeo dataset the models are trained on. The model architectures come from BEVFormer and CAPE
(Apache-2.0), and from GaussianLSS and Cross View Transformers (MIT). When using these checkpoints,
please cite RoboGeo, SimBEV and CARLA as listed in the
[dataset card](https://huggingface.co/datasets/shdragron/RoboGeo).