timb2001/plentiful-carla-camera-rigs

Benchmark: Plentiful Carla Camera Rigs

Camera-based perception systems for autonomous driving are typically developed and evaluated using fixed sensor rigs, while real-world vehicle fleets exhibit substantial variation in camera placement, orientation, field of view, and camera count. This mismatch introduces a cross-rig domain gap in which only the geometric observation process changes.

To study this effect under controlled conditions, we introduce Plentiful Carla Camera Rigs (PCCR), a 3D object detection benchmark that renders identical driving scenes under 14 systematically designed camera rigs. Using this benchmark, we analyze cross-rig transfer behavior of representative multi-view perception architectures and observe substantial performance shifts induced by geometric rig variation.

For more information, please visit our project page or GitHub repository. The latter contains benchmark utilities and the code for re-generating the data.

Structure

pccr/data/

The main data folder contains multi-camera images, LiDAR point clouds, annotations, and CAN bus data for all 14 camera rigs. It is organized as follows:

pccr/data/
├── R1.zip/                           # Baseline rig with 6 cameras + LiDAR
├── R1-c10.zip/                       # 10-camera variant
├── R1-c6.zip/                        # 6-camera variant  
├── R1-f.zip/                         # Front-facing variant
├── R1-r.zip/                         # Rear-facing variant
├── R1-t.zip/                         # Top-mounted variant
├── R2.zip/ to R9.zip/                # Additional rig configurations
│
└── [Each rig contains]:
    ├── *_infos_mini.pkl          # Annotation metadata (mmdet3d format)
    ├── *_infos_test.pkl
    ├── *_infos_train.pkl
    ├── *_infos_val.pkl
    │   # pkl files encode sample/annotation information for training with
    │   # mmdet3d and related frameworks
    │
    ├── samples/                  # Sensor data organized by type
    │   ├── CAM_BUMPER_FRONT/
    │   ├── CAM_BUMPER_BACK/
    │   ├── CAM_FENDER_FRONT_LEFT/
    │   ├── CAM_FENDER_FRONT_RIGHT/
    │   ├── CAM_FENDER_BACK_LEFT/
    │   ├── CAM_FENDER_BACK_RIGHT/
    │   ├── CAM_SIDE_LEFT/
    │   ├── CAM_SIDE_RIGHT/
    │   └── LIDAR_TOP/            # Point cloud frames
    │
    ├── v1.0-mini/                # nuScenes-format metadata
    │   ├── scene.json
    │   ├── sample.json
    │   ├── sample_data.json
    │   ├── sample_annotation.json
    │   ├── ego_pose.json
    │   ├── calibrated_sensor.json
    │   ├── sensor.json
    │   ├── instance.json
    │   ├── category.json
    │   ├── attribute.json
    │   ├── visibility.json
    │   ├── log.json
    │   ├── map.json
    │   └── additional_types.json
    │
    ├── v1.0-test/                # Same structure for different splits
    ├── v1.0-trainval/
    │
    ├── can_bus/                  # CAN bus simulator data
    └── logs/                     # Generation logs and metadata

Data Format Notes:

• Pkl files (*_infos_*.pkl): Serialized Python dictionaries containing annotations and metadata compatible with MMDetection3D training pipelines and related 3D detection frameworks
• Images: 1280×720 RGB images from several camera perspectives
• LiDAR: Point cloud data in standard nuScenes format
• Metadata: Full nuScenes-compatible JSON schema for reproducible benchmarking

pccr/trajectories_pruned/

High-level ego and object trajectories extracted from simulations, organized by data split:

pccr/trajectories_pruned/
├── scenes.json                   # Master scene configuration and metadata
│
├── mini/                         # Small validation split
│   ├── mini_01.h5
│   ├── mini_02.h5
│   ├── mini_03.h5
│   ├── mini_04.h5
│   └── mini_05.h5
│
├── test/                         # Test split
│   ├── test_01.h5
│   └── ...
│
└── trainval/                     # Training + validation split
    ├── trainval_01.h5
    └── ...

Trajectory Format:

• H5 files: HDF5 format containing serialized trajectory data (ego poses, object bounding boxes, velocities) for trajectory-based analysis and reproduction

Citation

Please use the following BibTeX entry to cite our work:

@inproceedings{bader2026crossrig,
    title     = {Understanding Cross-Rig Generalization in Automotive Perception: a Multi-Rig Benchmark and Rig Variation Metrics},
    author    = {Bader, Tim Alexander and Eberhardt, Tim Dieter and Dillitzer, Maximilian and Stork, Wilhelm},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year      = {2026},
    note      = {Project page: https://badertim.github.io/plentiful-carla-camera-rigs/}
}