PolySeg and Human Segmentation Benchmark for 2D Polygons

PolySeg is a web-based tool for collecting/visualizing human segmentation for 2D polygons for Dual-Space Decomposition of 2D Complex Shapes (Dude2D) project. PolySeg was developed by ZhonghuaXi. The server side is written in python with MySQL database and client side uses HTML5 canvas with SVG.

Data collected using PolySeg

We use PolySeg and Amazon Mechanical Turk to collect segmentation data. We use this data as the ground truth in the Dude2D project.

There were in total 142 people participated, and we obtained 3818 valid segmentations (more than 30000 cuts).
On average, each user spent around 90 seconds to segment a polygon, and made about 8.12 cuts for a polygon.

Download data
You can download all human segmentations polyseg_data.zip (1.0MB all polygon models are included)

Download PolySeg

You can download the source code (server side, db scheme and client side): polyseg_src.zip (1.5 MB all polygon models are included)

Try PolySeg

You can try PolySeg right now at http://turing.cs.gmu.edu:8635/

How it works

We deployed PolySeg to Amazon Mechanical Turk (MTurk) as Human Intelligence Tasks (HITs).
It can also run as a standalone version on any system with python and MySQL installed.

First, detailed instructions with both good and bad examples of polygon segmentation are shown to user to let them get familiar with the system.
And then, a polygon will be displayed on the screen, and user is asked to draw diagonals to segment the model.
Constraints were added to prevent user making segments outside or intersecting with the polygon or other existing cuts.

Admin users can visualize and rate each segmentation to filter out extremely bad results (participants did good job on average).
Average segmentation can be visualized by overlaying all segmentations with certain opacity.

How it looks

User Interface of PolySeg.

Visualize the average segmentations, commonly used diagonals are shown in darker color.

Rate for a single segmentation instance.

Locations of visitors to this page
Computer Science @ George Mason University