In our work we consider three types of data:
Below you can download our dataset and see examples of these images.
These images were rendered using 3D models of sedans, free HDR environment maps from the internet and MERL BRDFs using the Blinn-Phong model.
As real test images, we have recorded photos of six toy cars that were completely painted with a single car lacquer, placed in four different lighting conditions and photographed from five different views. Additionally, those real images were manually segmented from the background.
Objects and Reflectance maps preperation!
In order to provide an even more challenging test set, we collect an additional 32 car images from Internet search. Here we do not have access to groundtruth normals or reflectance maps, but the test provides a realistic test case for imaged-based editing methods. Again, we have manually segmented out the body of the car.