Unpack AnyShape Classes

Overview • How To Run • Explanation

Overview

In Supervisely, you have to define classes before labeling. The shape of class specifies instruments that are available in annotation interface. For example, if you need to label cars with polygonal tool, you have to create class Car with shape Polygon.

Supervisely supports the following shapes:

• Rectangle
• Polygon
• Bitmap
• Line (polyline)
• Point
• Keypoints (graphs) - sets of vertices connected by edges
• Cuboids (2D and 3D)
• AnyShape (will be explained below)
• ...

Example: objects of class Bitmap can be labeled with the following instruments: brush + eraser, pen, polygon, SmartTool. Whatever instrument is used, these objects always are saved as masks (raster).

What is AnyShape class?

Let's concider the following case: semantic segmentation of cars. You would like to label cars faster with SmartTool (NN that integrated to labeling interface and produces masks). If SmartTool produces inaccurate predictions you will label objects manually with Polygonal tool. So, you want to label objects of class car with absolutely different instruments: Polygonal tool (vector) and SmartTool (raster). How to do it?

First option: Create two separate classes: car_bitmap with shape Bitmap and car_polygon with shape Polygon. Thus SmartTool objects will be of class car_bitmap and polygonal objects will be of class car_polygon. The main drawback: this approach doubles up the number of classes and later (before NN training) you have to merge class pairs to a single one anyway. And what if you have tens or even hundreds of classes?

Second option: Create class car with shape AnyShape. It means that you can use all annotation instruments to label objects. The main drawback: your labeling team have to understand annotation requirements well. For example, if labelers are not restricted to specific instrument, you expect polygons but they can label objects with rectangles. ¯\_(ツ)_/¯

How To Run

Step 1: Run from context menu of project / dataset

Go to "Context Menu" (images project or dataset) -> "Run App" -> "Transform" -> "Unpack AnyShape Classes"

Step 2: Waiting until the task finishes

Once app is started, new task appear in workspace tasks. Monitor progress from "Tasks" list.

Explanation

• Result project name = original name + "(without AnyShape)" suffix

• Your data is safe: app creates new project with modified classes and objects. The original project remains unchanged

• All "AnyShape" classes will be unpacked to equivalent classes with strictly defined shapes. In our example the class car will be splited to classes car_polygon, car_bitmap, car_rectangle.

• Colors of new classes will be generated randomly