GTA-Crime
收藏资源简介:
GTA-Crime是一个专注于致命暴力事件的合成数据集,包括枪击和刺伤等场景,数据集从类似真实CCTV摄像头的多视角采集。为解决此类场景的稀有性,还发布了生成这些类型视频的框架。此外,还提出了一种基于Wasserstein对抗训练的片段级域适应策略,以弥合合成GTA-Crime特征与现实世界特征(如UCF-Crime)之间的差距。实验结果表明,GTA-Crime数据集可以显著提高致命暴力检测的准确性。
GTA-Crime is a synthetic dataset focused on fatal violent incidents, including scenarios such as shootings and stabbings. The dataset is collected from multiple perspectives similar to real CCTV cameras. To address the rarity of such scenarios, a framework for generating videos of these types is also released. Furthermore, a segment-level domain adaptation strategy based on Wasserstein adversarial training is proposed to bridge the gap between the features of the synthetic GTA-Crime dataset and real-world features such as UCF-Crime. Experimental results demonstrate that the GTA-Crime dataset can significantly improve the accuracy of fatal violence detection.
GTA-Crime 数据集概述
数据集简介
GTA-Crime是一个用于致命暴力检测的合成视频数据集,基于论文《GTA-Crime: Leveraging Synthetic Video and Feature-Level Domain Adaptation for Enhanced Fatal Violence Detection》构建。该数据集包含涉及刺杀和射击的致命场景视频。
数据内容
- 原始视频数据:包含通过扩展GTA5Event代码创建的合成致命场景视频
- 提取特征数据:
- I3D特征
- CLIP特征
数据获取
- GTA-Crime视频数据:https://drive.google.com/file/d/14mA5jgSIlfGdE6P-PgOr0bFCOyDSxGhc/view?usp=sharing
- GTA-Crime I3D特征:https://drive.google.com/file/d/14CQoSPS0iRwkTfA8AHcLwImruuUPAUdu/view?usp=sharing
- GTA-Crime CLIP特征:https://drive.google.com/file/d/1fK5B5tJ-dVDSsS8LLawF2hecnMXfBouI/view?usp=sharing
数据集构建
数据集构建说明详见:https://github.com/ta-ho/GTA-Crime/blob/main/construction.pdf
相关数据集
- UCF-Crime 10-crop I3D特征:https://github.com/Roc-Ng/DeepMIL
- UCF-Crime 10-crop CLIP特征:https://github.com/nwpu-zxr/VadCLIP
性能表现
数据集在多种暴力检测模型上的AUC性能表现:
| 方法 | UCF | UCF+GTA(无DA) | UCF+GTA(WGAN-GP) | UCF+GTA(CycleGAN) |
|---|---|---|---|---|
| RTFM | 85.43 | 84.98 | 87.27 | 87.28 |
| UR-DMU | 86.04 | 81.39 | 86.47 | 86.35 |
| MGFN | 82.55 | 79.37 | 83.64 | 84.35 |
| CLIP-TSA | 78.75 | 81.62 | 82.66 | 81.65 |
| VadCLIP | 73.59 | 74.60 | 74.79 | 76.84 |

- 1GTA-Crime: A Synthetic Dataset and Generation Framework for Fatal Violence Detection with Adversarial Snippet-Level Domain Adaptation韩国汉阳大学电气工程系、电子工程系、人工智能半导体工程系 · 2025年



