autoseg | Automatic Fast WebAssembly Image Segmentation

Are you looking to convert UDT annotations into png masks? Use udt-to-png

autoseg segments images from a list of points and polygons containing classifications. Autoseg was originally created by Severin Ibarluzea for usage with the Universal Data Tool.

Features

• Simple API
• WebWorker background computation
• Deterministic, suitable for mask compression
• Fast. WebAssembly compiled from state of the art C++ implementations of SLIC superpixeling and graph cut
• Fully compliant with the Universal Data Tool and UDT Format

Example Inputs -> Output

Input Image

Input Polygon

Output Mask

CLI

Make sure to format your file in the *.udt.json format.

npm install -g autoseg

# This will output a bunch of image pngs for the masks of a udt file
autoseg some_file.udt.json -o output-directory


autoseg --help

# Usage: autoseg path/to/dataset.udt.json -o output-masks-dir
#
# Options:
#   --help               Show help                                       [boolean]
#   --version            Show version number                             [boolean]
#   --use-sample-number  Use the number of the sample as the mask filename
#                       (mask0001.png, etc.)
#   --output-dir, -o     Output directory for masks                     [required]

API

const autoseg = require("autoseg") // OR require("autoseg/node")

await autoseg.loadImage({ data: imData, width: 320, height: 249 })

// Points/polygons are in the UDT Shape format
// https://github.com/UniversalDataTool/udt-format
const maskImageData = await autoseg.getMask([
  { regionType: "point", x: 50, y: 50, cls: 0 },
  { regionType: "point", x: 150, y: 150, cls: 1 },
])

// If you have a canvas, you can draw the image of the mask
canvasContext.putImageData(maskImageData, 0, 0)

// NOTE: nodejs doesn't have builtin support for ImageData, but you
// can use the returned data in a similar way, it's an object with
// { data: Uint8ClampedArray, width: number, height: number }

Configuring

await autoseg.setConfig({
  maxClusters: 1000,
  classNames: ["dog", "cat"],
  classColors: [0xffffffff, 0xff000000],
})