RoboGeo: Benchmarking Camera-Geometry Robustness in Multi-Camera BEV Perception

WACV 2027 · Evaluations & Datasets Track · Project page

Benchmark

RoboGeo is a benchmark for camera-geometry robustness in multi-camera BEV perception (3D object detection and BEV segmentation). It re-renders the same driving scenes in CARLA under changed camera geometries and provides two protocols: viewpoint perturbation, which rotates a single camera or all six cameras of the base rig, and cross-platform transfer, which deploys Sedan-trained models on the SUV and Bus camera layouts. Robustness is quantified by the mean Viewpoint Robustness Score (mVRS) and the Cross-Platform Transfer Score (CTS).

The paper evaluates six 3D object detection models (BEVDet, BEVDepth, BEVFormer, DFA3D, DETR3D, CAPE) and seven BEV segmentation models (LSS, GaussianLSS, BEVFormer, PointBeV, CVT, LaRa, DSPE) under one normal condition and three diagnostic conditions (EXT, IMG, CAL) in both protocols. It also tests extrinsic augmentation and two cross-domain generalization methods, PD-BEV and EAFormer. Code and evaluation harness: available soon.

What's inside

component	files	size	contents
metadata/carla_geobev_meta.tar.zst	1	1.5 GB	annotations in the nuScenes schema: v1.0-carla_{sedan,suv,bus} (training split, 220 scenes) and v1.0-carla_{sedan,suv,bus}_eval (evaluation split, 48 scenes), split/{train,val}.txt, detection info .pkls
platform_layouts/RGB-*.tar	18	273 GB	six-camera images for each platform: RGB-CAM_* = Sedan (base configuration), RGB-suv-CAM_* = SUV, RGB-bus-CAM_* = Bus
viewpoint_perturbation/RGB-*.tar (+viewpoint_metadata.json)	186	568 GB	viewpoint perturbation images, 50 scenes: 6 cameras × 31 settings (roll, pitch or yaw at ±4°, ±8°, ±12°, ±16°, ±20°, plus the base rig yaw0pitch0roll0); extrinsics in viewpoint_metadata.json
bev_segmentation_labels/*.tar	6	1.0 GB	BEV segmentation labels for the training split ({sedan,suv,bus}) and the evaluation split ({sedan,suv,bus}_eval) of each platform

Setup

mkdir -p carla_geobev carla_VR carla_geobev_labels

# metadata + images  → carla_geobev/
tar -I zstd -xf metadata/carla_geobev_meta.tar.zst -C carla_geobev
for t in platform_layouts/*.tar;  do tar -xf "$t" -C carla_geobev; done   # creates carla_geobev/sweeps/RGB-*/

# viewpoint perturbation images → carla_VR/
cp viewpoint_perturbation/viewpoint_metadata.json carla_VR/
for t in viewpoint_perturbation/*.tar; do tar -xf "$t" -C carla_VR; done          # creates carla_VR/sweeps/RGB-*_yaw*pitch*roll*/

# BEV segmentation labels → carla_geobev_labels/
for t in bev_segmentation_labels/*.tar; do tar -xf "$t" -C carla_geobev_labels; done   # creates carla_geobev_labels/{sedan,suv,bus}{,_eval}/

The extracted carla_geobev/ is what the benchmark configs point to as the data root (nuScenes-style layout). All paths inside the DBs and .pkls are relative: the DB JSONs and carla_infos_*.pkl use sweeps/RGB-*/... (relative to carla_geobev/), while {sedan,suv,bus}_infos_*.pkl follow the mmdet3d convention data/nuscenes/sweeps/... — symlink your extracted carla_geobev/ to <repo>/data/nuscenes for those. The pkls also reference sweeps/LIDAR/*.npz entries; LiDAR files are not part of this release (camera-only evaluation never loads them — see "Not included").

Frame-rate note (important)

carla_geobev frame N corresponds to carla_VR frame 2N. Annotations (DBs, pkls and BEV segmentation labels) cover the even carla_geobev frames: 79 per scene at 5 Hz, i.e. 17,380 training frames and 3,792 evaluation frames per platform. The platform_layouts shards also contain the odd frames (158 frames per scene), and the viewpoint_perturbation shards contain every even carla_VR frame (7,884 per shard). Only carla_VR frames 2N of annotated frames N (every fourth carla_VR frame, 3,792 per shard) have ground truth. Scenes scene_0248 and scene_0260 in viewpoint_perturbation belong to neither split.

Not included

LiDAR point clouds. They are part of the original SimBEV dataset.

BEV segmentation labels

bev_segmentation_labels/ holds the BEV vehicle segmentation labels used by every segmentation model in the benchmark, in the Cross View Transformers format: per scene, a JSON index of the annotated frames (image paths and camera parameters), and per frame a 200×200 BEV class map (±50 m, 0.5 m per cell), a visibility map and auxiliary vehicle targets.

License

RoboGeo is released under CC BY-NC-SA 4.0. It is derived from SimBEV, whose driving logs were replayed in CARLA to re-render the camera images, and is distributed with the permission of the SimBEV authors. It contains no original SimBEV files. When using RoboGeo, please also cite SimBEV and CARLA:

@article{mehr2025simbev,
  title   = {SimBEV: A Synthetic Multi-Task Multi-Sensor Driving Data Generation Tool and Dataset},
  author  = {Mehr, Goodarz and Eskandarian, Azim},
  journal = {arXiv preprint arXiv:2502.01894},
  year    = {2025}
}

@inproceedings{carla,
  title     = {{CARLA}: An Open Urban Driving Simulator},
  author    = {Dosovitskiy, Alexey and Ros, German and Codevilla, Felipe and Lopez, Antonio and Koltun, Vladlen},
  booktitle = {Proceedings of the 1st Annual Conference on Robot Learning (CoRL)},
  year      = {2017}
}

Citation

@inproceedings{robogeo2027,
  title     = {RoboGeo: Benchmarking Camera-Geometry Robustness in Multi-Camera BEV Perception},
  author    = {Jeongtae Kim and Giyong Moon and Chanyoung Lee and Soonmin Hwang},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}