hair-cam-anno

What This Skill Does

hair-cam-anno is a specialized AI agent skill designed for annotating surveillance camera footage to create high-quality datasets for training Vision-Language (VL) models. It streamlines the lifecycle of video-to-data preparation, covering three core stages: automated frame extraction, intelligent visual content analysis, and the generation of structured training manifests. By processing raw video files from local directories, this skill extracts key frames at specific intervals and pairs them with rich, structured metadata including environment descriptions, demographic information, behavioral analysis, and safety risk assessments. The final output is a standardized dataset.jsonl file, ready for immediate integration into fine-tuning pipelines for VL models. It ensures that surveillance data is not just stored, but transformed into semantically meaningful training examples.

Installation

You can install this skill directly via the OpenClaw CLI using the following command: clawhub install openclaw/skills/skills/aowind/sjht-cam-anno

Use Cases

• VL Model Fine-Tuning: Prepare high-quality, ground-truth annotated video datasets for training vision-language models to recognize security threats.
• Security Surveillance Analytics: Automate the description and classification of human behavior captured in standard security camera formats.
• Safety Compliance Auditing: Detect and document high-risk behaviors (e.g., falls, unauthorized entry) by generating structured risk assessments from video footage.
• Dataset Curation: Systematically process large batches of raw video recordings into uniform datasets for machine learning research.

Example Prompts

1. "Analyze the videos in /root/hair-cam/office-data and generate a dataset.jsonl file for fine-tuning our vision model."
2. "I have some new surveillance footage; please extract frames using the hair-cam-anno skill and help me write the descriptions."
3. "Can you process the security recordings in the local directory and flag any videos with high risk levels for the training dataset?"

Tips & Limitations

• Consistency: Always ensure the information inferred from filenames (such as occupancy counts or specific behaviors) aligns with the descriptive JSON to prevent label noise during model training.
• Framerate: The default frame extraction is set to 0.5 fps (1 frame every 2 seconds). Adjust this parameter if your specific security scenarios require higher temporal resolution for action recognition.
• Risk Assessment: Always reference the references/labels-reference.md file to ensure that your 'risk' level assignments follow the established schema for consistent dataset training.
• Data Privacy: Ensure that the video data residing in /root/hair-cam has been processed in accordance with your organization's privacy and data protection policies, especially when dealing with human subjects.