Data Generator

A virtual room in 3D space with scripts for changing interior.

INDUSTRY
Software & Hi-Tech

PROJECT TYPE
Mobile
Web

TECHNOLOGIES

unity3d
blender

As part of the development of an application for changing interior design, extra datasets were needed to train the neural network. It could help to better define the walls in real rooms. We decided to test the hypothesis that through generating of an artificial dataset would be possible to get good results in training a neural network. Then a lot of human resources would not be needed for wall marking.

The application uses several photos for training the neural network:

photo of the room itself;

a mask where there is only a white wall to define the walls in the room.

Two options were tested:

a set of photograph + mask (black and white, where the walls were colored in white, and the rest - were black)

a set of photograph + data (the position of the angles in pixels in the photo).

Task

Development of datasets for neural network training using Unity3D.

TECHNOLOGIES

unity3d
blender

Solution

Our Unity specialists created a virtual room in 3D space, where they arranged various interior objects and cameras. There we implemented scripts that change walls, wallpapers, interior items and create a randomly assembled room based on the position of the walls.

How are walls defined? After cameras set up, scripts start their rotation up to 360 degrees. When the camera rotates a certain number of degrees, depending on a given number of photos, the application takes a picture of what the camera sees now. Simultaneously the detected wall turns green, and then each green pixel changes to white, and all the others are painted black. That is how a black and white mask is created, where the wall is white.

To reproduce the process described above, we implemented the generation of a CSV file for each new photo. This file contains the coordinates of the corners of the walls.

As a result, it was decided that such an artificial generation of datasets could help in case when the neural network can not determine the atypical room or walls.