Learning Transparent Object Matting from a Single Image, CVPR 2018 (Spotlight),
Guanying Chen*, Kai Han*, Kwan-Yee K. Wong
(* indicates equal contribution)
As no off-the-shelf dataset for transparent object matting is available, and it is very tedious and difficult to produce a large real dataset with ground truth object masks, attenuation masks and refractive flow fields, we created a large-scale synthetic dataset by using POV-Ray to render images of synthetic transparent objects.
- POV-Ray 3.7
- Python 2.7
After installing POV-Ray, type the following command to render a transparent object in front of an image and get the ground truth object mask, attenuation mask and refractive flow field.
bash render.sh
We obtained the ground truth refractive flow field of an object by rendering it in front of a sequence of Gray-coded pattern.
We provide the 178K randomly generated POV-Ray objects in .inc
format for your usage.
bash download_objects.sh