Ourlabels - Web Based Community Annotation

A React.js website which can be used for annotating datasets (currently only rectangular annotations, but polygons to come)

All annotations are stored on a replicated, mongo database. A great resource for mongo databases is MongoDB Atlas. Raw annotation units are fractions proportionate to the image at whatever resolution is displayed (zoomed or not).

Annotation container

• Image component
  • A reusable component, but still needs a lot of work to truly be reusable
  • Uses react-konva to allow users to draw on an image and get sub-pixel specificity for coordinates
• Classification component
  • Controls classification selection
  • Controls zoom level
  • Controls label types

Projects

• A user with administrator role can create a project
• Projects can be of a predefined project type (currently: science, science video, civic, civic video)
• Projects may be public (open to anyone with an account) or private (open to users who have been approved by the project owner).
• Projects have
  • Labels (owners add these definitions for annotations - see labels)
  • Videos (Sequences of images) if a video type project
  • Images if a regular project
  • Owners
• Data from projects can be downloaded by:
  • Owners if a private project
  • Anyone if a public project

Videos

• Video projects are special and backend operators must have ffmpeg with the ability to process the types of videos that are accepted.
• Videos are limited to 15 MB and are split into individual frames (no matter what framerate the video is).
• Default accepted video types are: MPEG2, MPEG4, TS
• Annotation of video files is simplified by the presence of the Copy button on the Classification component. The Copy button allows a user to copy the previous frame's annotations, which the user can then alter for the current frame and save. It greatly reduces the cost of annotating.

Labels

• Labels for a project are determined by the project owner. Labels consist of a color, a label, and a description.
• Labels are displayed on the Classification component when annotating an image/frame.

Data

• Annotation data can be downloaded in two formats,
  • COCO JSON format
    • Occluded, Truncated, Difficult attributes are not retained.
  • PASCAL VOC XML format
    • Occluded, Truncated, Difficult attributes are retained.