---
task_categories:
  - object-detection
  - robotics
tags:
  - autonomous-driving
  - 3d-object-detection
  - robustness
  - point-cloud
  - multimodal
  - corruption-benchmark
pretty_name: ComboBench
---

# ComboBench

This repository hosts the dataset associated with the paper **Beyond Isolated Corruptions: Benchmarking 3D Robust Perception across Compositional Adversity Space**.

ComboBench is designed to benchmark robust 3D perception under compositional adversity. It contains corrupted sensor data generated by combining multiple perturbations, enabling the evaluation of perception systems under compound sensing degradations.

## Dataset Coverage

The release covers two benchmarks:

| Benchmark | Source Dataset | Released Severity Levels |
|---|---|---|
| KITTI-Combo | KITTI | 3 and 5+ |
| nuScenes-Combo | nuScenes | 3 and 5+ |

The severity notation follows the definitions in the paper. Severity levels describe corruption intensity, rather than the number of perturbations in a composition.

The broader ComboBench framework supports camera-based, LiDAR-only, and multimodal 3D perception research.

## Directory Structure

Both **KITTI-Combo** and **nuScenes-Combo** strictly preserve the directory hierarchy and file naming conventions of their respective source datasets (KITTI and nuScenes) for seamless integration with existing 3D perception codebases (e.g., MMDetection3D, OpenPCDet). 

> **Note:** Since ComboBench is designed for robustness evaluation and training-time data augmentation, the generated data **only includes the `training` and `validation` splits** (usually organized under `training/` for KITTI and `v1.0-trainval` for nuScenes), omitting the official `testing` splits. For uninfected metadata, annotations, and calibration files, you can either symlink them from the clean source datasets or place the corrupted sensor folders directly into your existing dataset root.

<table>
<tr>
<th width="50%">KITTI-Combo Structure</th>
<th width="50%">nuScenes-Combo Structure</th>
</tr>
<tr>
<td valign="top">

```text
KITTI-Combo/
└── <corruption_type>_<severity>/
    ├── ImageSets/
    │   ├── train.txt
    │   ├── val.txt
    │   └── trainval.txt
    └── training/
        ├── calib/
        │   └── *.txt
        ├── image_2/          # Corrupted RGB images
        │   └── *.png
        ├── label_2/
        │   └── *.txt
        └── velodyne/         # Corrupted LiDAR point clouds
            └── *.bin
```

</td>
<td valign="top">

```text
nuScenes-Combo/
└── <corruption_type>_<severity>/
    ├── maps/
    ├── samples/              # Keyframes (train & val)
    │   ├── CAM_FRONT/        # Corrupted multi-view images
    │   ├── CAM_FRONT_LEFT/
    │   ├── CAM_FRONT_RIGHT/
    │   ├── CAM_BACK/
    │   ├── CAM_BACK_LEFT/
    │   ├── CAM_BACK_RIGHT/
    │   └── LIDAR_TOP/        # Corrupted LiDAR (.pcd.bin)
    ├── sweeps/               # Intermediate frames (if applicable)
    │   ├── CAM_*/
    │   └── LIDAR_TOP/
    └── v1.0-trainval/        # Metadata & annotations
        └── *.json
```

</td>
</tr>
</table>

## Intended Uses

- Evaluating 3D object detection robustness under combined corruptions.
- Comparing model performance across corruption compositions and severity levels.
- Studying sensor degradation and cross-modal robustness.
- Developing and evaluating methods for robust 3D perception.

## Access and Usage

Please see the **Files and versions** tab for the currently available data files. Refer to the paper and code repository for the corruption-generation methodology and evaluation protocols.

- **Code:** https://github.com/buaa-linx/ComboBench
- **Project page:** https://buaa-linx.github.io/ComboBench/

## Data Sources

KITTI-Combo and nuScenes-Combo are derived from KITTI and nuScenes, respectively. Please acknowledge the original datasets alongside ComboBench when using this release.

## Citation

If you use this dataset in your research, please cite **Beyond Isolated Corruptions: Benchmarking 3D Robust Perception across Compositional Adversity Space**, together with the relevant source dataset.