Sparse, Smart Contours to Represent and Edit Images

Supplementary Material

 

We show reconstruction results on "VGG-Faces", "Caltech-CUB" and "Stanford Dogs" datasets. For each of the datasets, we trained a new model and show 30 randomly sampled images from the test set. We also show results on CelebA where the model was trained on VGG dataset. For each image, we show our reconstruction using three types of features: gradients, color (RGB) and learned features (see Section 4 in the paper). For VGG and CUB datasets we show a comparison to Pix2Pix (trained using their implementaiton on our data). All results are on images from test sets.

 

		"VGG Faces" (256 x 256) Source: http://www.robots.ox.ac.uk/~vgg/data/vgg_face/ 30227 Training Images Sampled reconstruction results
		"VGG Faces" (512 x 512) Source: http://www.robots.ox.ac.uk/~vgg/data/vgg_face/ 30227 Training Images Sampled reconstruction results
		"Caltech-CUB" 8855 Training images. Source: http://www.vision.caltech.edu/visipedia/CUB-200.html Sampled reconstruction results
		"Stanford Dogs" 12000 images. Source: http://vision.stanford.edu/aditya86/ImageNetDogs/ Sampled reconstruction results
		"CelebA" Results on CelebA: Model was trained on vgg dataset. Sampled reconstruction results