-
Notifications
You must be signed in to change notification settings - Fork 464
Home
COCO Annotator is a web-based image annotation tool designed for versatility and ease of use for efficiently label images to create training data for image localization and object detection. It provides many distinct features including the ability to label an image segment (or part of a segment), track object instances, labeling objects with disconnected visible parts, efficiently storing and export annotations in the well-know COCO format. Above all, COCO Annotator automates the process of annotating images to allow rapid growth of datasets.
Computer vision systems have become significantly more efficient at tasks such as object detection, tracking, classification, and segmentation. Many of those systems utilize supervised learning techniques and train on large datasets of annotated images. Hence efforts have now shifted to produced reliable datasets for training many of these computer vision systems. As these datasets grow, annotating the images becomes increasingly more challenging and the efficiency of the annotation tool used becomes more essential. Unfortunately, very few annotation tools address this need for efficiency and versatility.
Several annotation tools are currently available, with most applications as a desktop installation. Once installed, users can manually define regions in an image and creating a textual description. Generally, objects can be marked by a bounding box, either directly, through a masking tool, or by marking points to define the containing area. COCO Annotator allows users to annotate images using free-form curves or polygons and provides many additional features were other annotations tool fall short.
- Directly export to COCO format
- Segmentation of objects
- Useful API endpoints to analyze data
- Import datasets already annotated in COCO format
- Annotated disconnected objects as a single instance
- Labeling image segments with any number of labels simultaneously
- Allow custom metadata for each instance or object
- Magic wand/select tool
- Generate datasets using google images
For cluttered images such as that in Figure 1, the image tends to be dense with many objects. In this figure, we can see numerous cars were the process to annotate each one becomes time-consuming. When tens of thousands such images need to be ground-truthed, an efficient annotation tool becomes more pressing. For starters, some type of semi-automatic marking of the image segments may be helpful in speeding up the annotation process.
Figure 1: Traffic on an highway in Delhi
This problem is solved through the use of two different methods. A tool called Magic Wand uses the flood fill algorithm to create a selection of pixels similar in color and shade. The second method allows users to configure external API call to a semi-trained model which then applies annotations return from the request.
Annotated objects are not required to be composed of continuous segments. For example, if a car is partially blocked by a tree. The disjoint visible parts can be annotated as part of a single car without including the tree.
Figure 2: Annotation of a single car object with visual disconnected parts