Datasets:

BodyCam-VQA
/

BWC-VideoText-359

task_categories:
  - visual-question-answering
language:
  - en
tags:
  - law-enforcement
  - legal
  - body-worn-camera
  - surveillance
  - copa
  - ground-truth
pretty_name: BodyCam-VQA (BWC-VideoText-359)
size_categories:
  - n<1K

BodyCam-VQA: BWC-VideoText-359

Dataset Summary

BodyCam-VQA (BWC-VideoText-359) is a specialized multimodal dataset designed for research in legal transparency, forensic linguistics, and automated video analysis. It consists of 359 one-minute video segments sourced from the Civilian Office of Police Accountability (COPA) of Chicago.

This dataset is unique in its focus on Body-Worn Camera (BWC) footage, which presents distinct challenges such as occlusions, rapid motion, varied lighting, and high-ambient-noise audio environments.

Dataset Structure

Data Statistics

The dataset is split into training and evaluation subsets. The evaluation set is accompanied by a human-verified "gold standard" ground truth file for benchmarking.

Component	Split	Count	Format
Videos	Train	288	`.mp4`
Transcripts (WhisperX)	Train	288	`.json`
Videos	Evaluation	71	`.mp4`
Transcripts (WhisperX)	Evaluation	71	`.json`
Ground Truth (Human)	Evaluation	1	`.json`

File Organization & Naming

Video Files: Located in train_videos.zip and eval_videos.zip. Filename format: Video[index].mp4.
Machine Transcripts: Located in train_transcripts.zip and eval_transcripts.zip. These can be mapped to videos via the index.
Evaluation Ground Truth: The file human_annotated_ground_truth_eval_set.json contains the human-evaluated labels for the 71 evaluation videos, serving as the definitive benchmark for model performance.

Data Generation & Quality Control

Automated Transcription

Initial transcripts were generated using WhisperX. This tool was selected for its robust performance in noisy environments and its ability to provide precise word-level timestamps, essential for temporal alignment in video-text tasks.

Human Annotation (Evaluation Set)

To ensure the reliability of the evaluation benchmark, the 71-video evaluation subset was manually reviewed and annotated. This Ground Truth file corrects potential machine-transcription errors and provides verified labels for Visual Question Answering (VQA) tasks, ensuring that model assessments reflect real-world accuracy.

Usage & Benchmarking

Benchmarking Protocol

Researchers should use the train split for model fine-tuning and the eval split for testing. Performance should be measured against the human_annotated_ground_truth_eval_set.json to ensure results are compared against a human-verified baseline.

Intended Use

Visual Question Answering (VQA): Captioning video events via natural language queries.
Legal & Forensic Analysis: Training models to assist in the review of law enforcement public records.
Multimodal Robustness: Testing AI performance on real-world footage.

Ethical Considerations & Licensing

Ethics: This dataset contains real-world law enforcement interactions. While these are public records released in the interest of transparency, researchers are expected to use this data ethically and responsibly. Avoid using this dataset for person identification or unauthorized surveillance applications.

License: This data is sourced from public records provided by the City of Chicago. It is subject to the terms and conditions of the Civilian Office of Police Accountability (COPA) public disclosure policies.