Notate ML 1.03

Preparing input image datasets is an essential and often complex aspect of any Machine Learning project. Notate ML harnesses the capabilities of Apple's mobile devices to streamline this process and enhance the quality of training data for object detection models.

Features

• Create a dataset with labels via typing, scanning, or voice input.
• Capture new images or import existing ones from your photo library.
• Edit images by cropping, drawing bounding boxes, and tagging objects of interest.
• Export images and annotations for training with frameworks such as YOLO, Apple Create ML, or Google Auto ML.

Creating a Dataset

• Initiate a new dataset by assigning it a unique name.
• Add labels through various methods:
  • Individually using the keyboard.
  • Pasting a comma-separated list from external sources.
  • Scanning text from physical media.
  • Utilizing voice input on the device.
• Edit or delete labels by swiping left on each entry; note that only unused labels may be removed.

Importing Images

• Access the dataset to import photos (up to 50 at once) from your Photo Library, using Apple's photo picker feature for selection.

Capturing Images from the Camera

• While viewing images in a dataset, select the "Capture" button to take new photos with the camera.
• The app will request permission to access the camera; this can be granted later via "Settings" if declined initially.

Annotating the Image

• Select an image from the dataset to annotate it.
  • Zooming and Cropping:
    • Pinch with two fingers to zoom in or out, and drag to pan across the image.
    • Crop the image only when there are no bounding boxes present; double-tap to reset zoom and recenter.
    • The dimensions of cropped images are constantly displayed above the image for reference.
  • Tagging:
    • Draw bounding boxes around areas of interest using one finger or a stylus.
    • Select existing annotations by long-pressing on a box, and assign labels using a picker.
    • Tap outside a selected box to deselect, or use the red Delete button for removal.
  • To finalize annotations, tap the "Done" button after editing, or use "Reset" to revert changes made before completion.

Exporting the Dataset

• Select the "Export" button while browsing dataset images to choose from various export options.
• Diverse schemas are available for exporting datasets tailored for different object detection training frameworks:
  • The YOLO schema formats files for YOLO framework usage.
  • The Create ML schema produces files for Apple's Create ML application input.
  • The Auto ML schema generates files compatible with Google's Auto ML service.
• After reviewing options, confirm your choice by tapping "Export" to generate necessary data files and select a sharing method (e.g., Airdrop, Files) for exporting these files.

Deleting Content

• Images can be deleted by swiping left in the dataset view, while entire datasets can also be removed under similar actions; confirmation will be required if images are present in the dataset.

Tips for Effective Use

• Crop images prior to tagging to enhance annotation accuracy.
• A stylus is recommended for more precise bounding box creation.
• Create separate datasets for training and validation purposes.
• Limit dataset sizes to under 1000 images each to aid usability and performance.
• Regularly export and delete any unused datasets to conserve storage space on your device.