data

📊 Data Directory

This directory contains all data files, datasets, and evaluation results in an organized structure.

📁 Directory Structure

data/
├── datasets/           # Ground truth and training datasets
│   ├── ml_dataset.csv         # Main ground truth dataset
│   └── cooking_actions_log.csv # Recording session logs
├── sample_videos/      # Sample/test videos for evaluation
│   └── README.md              # Video usage instructions
└── results/            # Evaluation outputs and analysis results
    ├── classification_result.txt        # Example log file
    ├── classification_result_timeline.log  # Example timeline
    ├── action_timeline.png             # Visualization
    └── cooking_actions_recording.mp4   # Example recording

📋 Dataset Descriptions

datasets/ml_dataset.csv

Ground truth dataset for evaluation

Contains labeled action segments for performance comparison:

• action_label: Specific action (add-food, remove-lid, etc.)
• start_time_seconds: Action start time
• end_time_seconds: Action end time
• duration_seconds: Action duration
• video_filename: Source video file

Usage:

cd evaluation/
python3 extract_and_align_classifier.py logs.txt
# Automatically loads ../data/datasets/ml_dataset.csv for comparison

datasets/cooking_actions_log.csv

Recording session metadata

Generated by interactive recording sessions:

• Timestamped action labels
• Action categories and descriptions
• Video filenames and metadata
• Session information

🎥 Sample Videos

Place your evaluation videos in sample_videos/ directory:

# Copy your videos
cp /path/to/your/videos/*.mp4 data/sample_videos/

# Run evaluation
cd evaluation/
python3 automate_evaluation.py --video-dir ../data/sample_videos

Supported formats: .mp4, .avi, .mov, .mkv

📊 Results Directory

Evaluation outputs are automatically saved here:

Per-Video Results

• videoname_timestamp.log - Raw viam-server logs
• videoname_timestamp.analysis.txt - Detailed performance analysis
• videoname_timestamp_timeline.log - Extracted classification timeline

Summary Reports

• evaluation_results_timestamp.json - Complete evaluation summary
• Cross-video performance metrics
• Aggregate statistics and recommendations

Example Results Structure

results/evaluation_20250611_120000/
├── video1_20250611_120001.log
├── video1_20250611_120001.analysis.txt
├── video2_20250611_120035.log
├── video2_20250611_120035.analysis.txt
└── evaluation_results_20250611_120000.json

🔄 Data Flow

Recording → sample_videos/ → Evaluation → results/
                ↓                ↓
            datasets/        Comparison with
         (ground truth)     ground truth data

For Recording:

1. Use recording/ tools to capture videos
2. Videos saved to sample_videos/ for evaluation
3. Action logs saved to datasets/ for ground truth

For Evaluation:

1. Videos from sample_videos/ processed automatically
2. Results compared against datasets/ml_dataset.csv
3. Outputs saved to results/ with timestamps

🛠️ Data Management

Adding New Videos

# Copy videos to sample directory
cp /path/to/new/videos/*.mp4 data/sample_videos/

# Or specify custom directory in evaluation
python3 automate_evaluation.py --video-dir /custom/path/to/videos

Updating Ground Truth

# Update the ground truth dataset
cp new_ml_dataset.csv data/datasets/ml_dataset.csv

# Evaluation will automatically use updated ground truth

Cleaning Results

# Remove old evaluation results (optional)
rm -rf data/results/evaluation_older_timestamp/

# Keep recent results for comparison
ls -la data/results/

📈 Integration

With Recording Tools

# Record videos directly to data directory
cd recording/
OUTPUT_DIR="../data/sample_videos" python3 record_rgb_interactive.py

With Evaluation Pipeline

# Automated evaluation of all data
cd evaluation/
python3 automate_evaluation.py \
    --video-dir ../data/sample_videos \
    --output-dir ../data/results/evaluation_$(date +%Y%m%d_%H%M%S)

With External Tools

import json
import pandas as pd

# Load evaluation results
with open('data/results/evaluation_results_timestamp.json') as f:
    results = json.load(f)

# Load ground truth
gt_data = pd.read_csv('data/datasets/ml_dataset.csv')

# Integrate with your analysis pipeline...

📋 File Formats

Ground Truth CSV

video_filename,action_label,start_time_seconds,end_time_seconds,duration_seconds
video1.mp4,add-food,15.2,18.4,3.2
video1.mp4,flip,25.1,28.3,3.2

Evaluation Results JSON

{
  "timestamp": "2025-06-11T12:00:00",
  "machine_id": "de6836af-05f6-4ff4-a067-8d18ac0f6495",
  "videos": [...],
  "summary": {
    "total_detections": 25,
    "overall_acceptance_rate": 85.2,
    "avg_confidence_across_videos": 78.5
  }
}