isp-uv-es/IPL-Cityscapes-Grayscale-Illuminants

IPL-Cityscapes-Grayscale-Illuminants dataset

Illuminant modified dataset version of the grayscale famous autonomous driving semantic segmentation Cityscapes dataset.

Dataset generation

For each image, we first convert it to grayscale. Then, we generate a flat (constant) light spectrum and compute the pixel reflectances that obtain the RGM pixel values. Once we have the pixel reflectances, we generate different light spectrums of different dominant wavelengths (colors) and saturations and compute the new modified images. We apply the method to 100 Cityscapes images. For each original image we generate 120 modified versions (20 colors and 6 saturations).

Dataset structure

The originalimages folder contains the original 100 grayscale Cityscapes images. The modifiedimages folder have 100 subfolders from image1 to image100. Each subfolder has the 120 modified images for each original image. Modified images are named as imagenameXtonoTsatS.png, where X is the image index from 1 to 100, T is the color index from 1 to 20 and S is the saturation index from 1 to 6. For the labels see the labels folder in IPL-CityscapesIlluminants-dataset, which contains the corresponding ground truth segmentation masks. The complete dataset (original grayscale + modified images) has a size of 30 Gb.