>
About me
Projects
Curriculum Vitae
LinkedIn
GitHub
Description
IllumiSet is a dataset containing more than 15k RGB-D images of synthetic indoor scenes. Each RGB-D image is accompanied by a normal image (containing surface normals for each pixel in camera-space), albedo (absolute colour of a surface),
and fine-grained shading images. These shading images include direct shading, indirect shading and glossy reflections. Additionally, for each light source active in an image,
the location in 3D-space is available. Other datasets of intrinsic images do not or partially include light source locations. The light source locations, however, allows the training if
IllumiNet, a deep convolutional neural network that recovers light source locations and intensities such that these can be used for
object insertion. However, not all illuminants have an explicit light source location, such as directional lights or environment maps. As a result, all light sources in the scene are point lights, set up in such a way that realistic lighting
is created.
Creation
IllumiSet is created by manually decorating and adjusting empty room layouts in Blender. To create a diverse enough dataset, the scene is filled with manually-placed cameras
and light sources. For each render, a camera and a set of light sources are selected and activated. The properties of the light source, including the colour, intensity and radius, are randomized. In total, IllumiSet is created using
11 indoor scenes shown below.
