| --- |
| license: cc-by-sa-4.0 |
| --- |
| # Dataset Card for SafeMVDrive |
|
|
| ## Dataset Description |
|
|
| ### Overview |
|
|
| The SafeMVDrive dataset is a collection of realism, multi-view safety-critical driving scenarios generated by the *SafeMVDrive* framework. The scenarios are generated starting from samples randomly selected from the nuScenes validation set, where an adversarial vehicle performs aggressive maneuvers to crash into the ego vehicle, and the ego vehicle reacts in time to avoid the collision. The dataset can be used to train and evaluate end-to-end autonomous driving models on their ability to handle safety-critical situations. |
|
|
| - **Curated by**: Jiawei Zhou, Linye Lyu, Zhuotao Tian, Cheng Zhuo, Yu Li |
| - **Affiliation**: Harbin Institute of Technology, Shenzhen; Zhejiang University |
| - **License**: CC-BY-SA-4.0 |
| - **Project Page**: [SafeMVDrive](https://zhoujiawei3.github.io/SafeMVDrive/) |
|
|
| ### Dataset Structure |
|
|
| The dataset consists of: |
| - **41 scenes** each 9 seconds long |
| - **Multiview video data** from 6 camera perspectives |
| - **3D bounding box annotations** Vehicles in the scene |
|
|
|
|
| ### Usage |
|
|
| The SafeMVDrive dataset is in nuScenes format. Please follow instructions to use SafeMVDrive dataset to evaluate the end-to-end driving model，Uniad， in [Eval Uniad](https://github.com/zhoujiawei3/SafeMVDrive#5-evaluating-end-to-end-autonomous-driving-models-on-the-generated-adversarial-dataset). |
|
|
| ## Creation Process |
|
|
| ### Source Data |
| - Built upon the nuScenes validation set (250 randomly selected samples) |
| - Uses nuScenes' original sensor data and annotations as foundation |
|
|
| ### Generation Pipeline |
| 1. **Adversarial Vehicle Selection**: Selects the most threatening vehicle based on multi-view visual input using a GRPO-finetuned Vision-Language Model (VLM). |
|
|
| 2. **Two-Stage Trajectory Generation**: |
| - *Collision Simulation*: Creates aggressive trajectories that cause the adversarial vehicle to collide with the ego vehicle. |
| - *Evasion Refinement*: Converts collision trajectories into realistic evasive maneuvers that avoid collision while retaining safety-critical properties. |
|
|
| 3. **Video Synthesis**: Produces high-fidelity, long-horizon, multi-view driving videos with UniMLVG diffusion video generator. |
|
|
| ### Filtering |
| Scenarios are filtered to ensure: |
| - During the collision stage, the adversarial vehicle collides with the ego vehicle, without entering non-drivable areas or colliding with any other vehicles beforehand |
| - During the evasion stage, the ego vehicle successfully avoids the adversarial vehicle, without colliding with any other vehicles or entering non-drivable areas |
|
|
| ## Intended Use |
|
|
| - **Robustness evaluation** of autonomous driving systems |
| - **Stress-testing** end-to-end AD models (e.g., UniAD, VAD, SparseDrive, DiffusionDrive) |
| - **Training end-to-end driving models** to learn evasive behaviors in safety-critical scenarios |
|
|
| ### Limitations |
| - Although guidance signals are used to generate annotations, the framework lacks a mechanism to discard outdated or irrelevant ones—for example, vehicles that have already exited the ego vehicle’s field of view. |
| - Open-loop evaluation (no reactive ego agent) |
| - Rendering artifacts compared to real sensor data |
|
|
|
|
| ### Privacy |
| - Based on nuScenes data which has already undergone anonymization |
| - No additional privacy concerns introduced by generation process |
