Data Preparation & Annotation

Supported Data Formats

Images: JPG, PNG (max 100MB each, recommended 640×640 to 1920×1080)

Annotation format: YOLO normalized coordinates (box center x, box center y, width, height)

Uploading Images

Web Drag-and-Drop Upload

Drag image folders directly into the upload area. Batch upload supported with real-time progress.

Mobile Photo Capture

For on-site training data collection using your phone:

  1. Click "Mobile Capture" on the dataset detail page
  2. A QR code is generated
  3. Scan the QR code with your phone to open the capture page
  4. Take photos of target objects
  5. Photos are automatically uploaded to the dataset

No app installation needed — just your phone's browser via QR code.

Annotation Best Practices

  1. Tight fit: Draw boxes that closely fit the target object — avoid excessive blank space
  2. Occlusion handling: Annotate the visible portion of occluded objects as well
  3. Quantity requirement: At least 50 instances per class, 200+ recommended
  4. Scene coverage: Cover different lighting, angles, and background variations
  5. Save promptly: Always click "Save" after completing annotations

Version Snapshots

A snapshot is a "frozen version" of your dataset at a point in time. Training must be based on a snapshot to ensure reproducibility.

  1. Click "Create Snapshot" on the dataset detail page
  2. Enter a version description (e.g., "v2.0 added 200 nighttime scenes")
  3. The platform auto-exports all annotations to YOLO training format

Dataset Management

On the dataset detail page you can:

  • Filter by annotation status (All / Annotated / Unannotated)
  • Delete unwanted images
  • View annotation progress (annotated count / total images)

Next