Research map / Video / Video generation
SMD: Multi-view Safety-Critical Driving Video Generation in the Real-world Domain
Jiawei Zhou, Linye Lyu, Zhuotao Tian, Cheng Zhuo, Yu Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add evasion stage after collision stage (two-stage trajectory generation) (ablation) | SMD · NuScenes val 250 samples | Natural score: 0.33 -> 0.57 | Improved |
| add evasion stage after collision stage (ablation) | SMD · NuScenes val 250 samples | Sample-level collision rate avg (%): 29.22 -> 15.23 | Hurt |
| use two-stage evasion trajectories instead of naive collision trajectories | SMD · NuScenes val 250 samples | Natural score: 0.21 -> 0.56 | Improved |
| use SMD trajectories instead of original NuScenes trajectories | SMD · planner stress test | Sample-level collision rate avg (%): 1.04 -> 15.23 | Improved |
| replace heuristic selection with GRPO-fine-tuned VLM selector (ablation) | SMD · adversarial vehicle selection | F1: 0.65 -> 0.71 | Improved |
| add GRPO fine-tuning to Qwen2.5VL-7B (ablation) | VLM selector · adversarial vehicle selection | F1: 0.49 -> 0.71 | Improved |
| replace SFT with GRPO for VLM fine-tuning (ablation) | VLM selector · adversarial vehicle selection | F1: 0.66 -> 0.71 | Improved |
| scale base VLM from 3B to 7B (no fine-tuning) (ablation) | Qwen2.5VL · adversarial vehicle selection | F1: 0.45 -> 0.49 | Improved |
| add SMD-generated videos to planner fine-tuning data | DiffusionDrive · SMD validation set | Sample-level collision rate avg (%): 16.64 -> 7.68 | Improved |
| add SMD-generated videos to planner fine-tuning data | DiffusionDrive · NuScenes val | Sample-level collision rate avg (%): 0.18 -> 0.13 | Improved |
| fine-tune on NuScenes only | DiffusionDrive · SMD validation set | Sample-level collision rate avg (%): 16.64 -> 19.42 | Hurt |
| fine-tune on NuScenes only | DiffusionDrive · NuScenes val | Sample-level collision rate avg (%): 0.18 -> 0.17 | No clear effect |
Closest papers by meaning
- Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving (CVPR 2024 · similarity 0.90)
- DiffScene: Diffusion-Based Safety-Critical Scenario Generation for Autonomous Vehicles (AAAI 2025 · similarity 0.88)
- DriveDreamer-2: LLM-Enhanced World Models for Diverse Driving Video Generation (AAAI 2025 · similarity 0.87)
- SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models (ICML 2025 · similarity 0.87)
- DriveScape: High-Resolution Driving Video Generation by Multi-View Feature Fusion (CVPR 2025 · similarity 0.86)
- DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
