| --- |
| license: cc-by-4.0 |
| task_categories: |
| - image-text-to-text |
| - image-to-text |
| - text-classification |
| language: |
| - en |
| tags: |
| - Image |
| - Captioning |
| - Violence |
| - Detection |
| - Vision |
| - AI |
| - Data |
| - Sceince |
| - Deep |
| - Learning |
| pretty_name: 'Violence-Cap: The First Frame-Level Captioning Dataset for Violence Research' |
| size_categories: |
| - 10K<n<100K |
| --- |
| # Violence-Cap Dataset |
|
|
|  |
|
|
| **Violence-Cap** is a video frame captioning dataset layered over the Real Life Violence Situations (RLVS) dataset. It is designed to provide granular, frame-level context for video violence analysis. |
|
|
| Each record in the dataset pairs an individual video frame with: |
| 1. A descriptive text caption. |
| 2. An independent, frame-level binary violence label. |
|
|
| ## Motivation & Key Findings |
| Video-level labels often fail to capture the temporal nuances of an event. In the creation of this dataset, 5,688 frames were extracted from videos globally labeled as *violent*. Upon human review, **1,243 frames (21.9%) were identified as *non-violent***. This highlights the critical necessity of per-frame annotation when training precise violence recognition and captioning models. |
|
|
| ## Dataset Statistics & Splits |
| The dataset features pre-defined, video-level **train, validation, and test splits** to ensure no data leakage between frames of the same video during model evaluation. |
|
|
| * **Curated Set:** 10,907 frame–caption–label triples across 1,991 videos. |
| * **Full Set:** 10,952 frame–caption–label triples across all 2,000 videos. |
|
|
| ## Files and Annotations |
| The annotations are distributed across two primary CSV files: |
|
|
| | File Name | Rows (Triples) | Video Count | Description | |
| | :--- | :--- | :--- | :--- | |
| | `curated_set.csv` | 10,907 | 1,991 | The thoroughly reviewed and cleaned dataset recommended for training and evaluation. | |
| | `full_set.csv` | 10,952 | 2,000 | The complete, unpruned dataset spanning all original RLVS videos. | |
|
|
| ## License |
| This dataset is released under the **Creative Commons Attribution 4.0 International ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/))** license. |
|
|
| ## Related Work |
| Violence-Cap is built upon the RLVS dataset. |
|
|
| > Soliman, M. M., Hussein Kamal, M., Abd El-Massih Nashed, M., Mostafa, Y. M., Safwat Chawky, B., & Khattab, D. (2019). *Violence recognition from videos using deep learning techniques.* 2019 Ninth International Conference on Intelligent Computing and Information Systems (ICICIS), 80–85. |
