data_collection_plan.md

Data Collection Strategy for Object Detection with Transformers

Objective

The objective of this data collection strategy is to gather image data suitable for object detection tasks within the PikasuBirdAi project, utilizing Hugging Face transformer models and the YOLO object detection framework with COCO dataset names.

Transformer Models

• DETR (DEtection TRansformer) for end-to-end object detection
• ViT (Vision Transformer) for image classification and feature extraction
• YOLOS (You Only Look at One Sequence) for object detection

Data Types

• High-resolution images of birds in various natural habitats
• Images with multiple birds of different species
• Images capturing birds in flight, perched, and in nests
• Varied lighting conditions (daylight, dusk, dawn)
• Different weather conditions (clear, cloudy, rainy)

Data Sources

• Public datasets: COCO dataset, iNaturalist, NABirds
• Flickr API for user-generated bird images
• Web scraping from reputable bird-watching websites and forums

Preprocessing Steps

1. Image resizing: Standardize to 640x640 pixels for YOLO compatibility
2. Normalization: Scale pixel values to [0, 1]
3. Data augmentation:
   • Random horizontal flips
   • Random rotations (±15 degrees)
   • Random brightness and contrast adjustments
4. Annotation conversion: Convert existing annotations to YOLO format

Data Storage

• Format: JPEG for images, TXT files for annotations
• Structure:

  data/
  ├── images/
  │   ├── train/
  │   ├── val/
  │   └── test/
  └── labels/
      ├── train/
      ├── val/
      └── test/
  

Integration with YOLO and COCO

• Map bird species to COCO class IDs
• Create a custom coco_bird_names.txt file with bird species names
• Use YOLO's dataset loading utilities to integrate with COCO format
• Implement a custom dataset class that combines COCO annotations with our bird-specific data

Documentation

• Maintain a README.md file in the data directory
• Document data collection sources, preprocessing steps, and usage instructions
• Keep a version history of the dataset
• Include scripts for data preprocessing and format conversion