The repository contains images used in subjective comparisons of automatic image inpainting methods vs. human artists. The study description is available at Towards Data Science.
The 33 photo patches used in the study are available in ground_truth directory. Distorted images given to inpainting mathods and human artists are located in input directory.
The results of automatic inpainting methods are located in results/automatic_methods directory. The file names abide the following pattern: results/automatic_methods/{TEST_IMAGE_NAME}/{METHOD_NAME}.png
.
Deep learning methods names:
deep_image_prior
: Deep Image Prior (Ulyanov, Vedaldi, and Lempitsky, 2017)globally_and_locally_consistent
: Globally and Locally Consistent Image Completion (Iizuka, Simo-Serra, and Ishikawa, 2017)high_res_neural_inpainting
: High-Resolution Image Inpainting (Yang et al., 2017)shift_net
: Shift-Net (Yan et al., 2018)generative_inpainting_release_*_256
: Generative Image Inpainting With Contextual Attention (Yu et al., 2018) — this method appears twice in our results because we tested two versions, each trained on a different data set (ImageNet and Places2)irregular_holes
: Image Inpainting for Irregular Holes Using Partial Convolutions (Liu et al., 2018)
Conventional inpainting methods:
criminisi_*
: Exemplar-Based Image Inpainting (Criminisi, Pérez, and Toyama, 2004). The number in the suffix encodes patch size (9 or 13).patch_shift_stats
: Statistics of Patch Offsets for Image Completion (He and Sun, 2012)photoshop_cs_5
: Content-Aware Fill in Adobe Photoshop CS5
The images drawn by human artists are located in results/human_artists folder. The file names abide the following pattern: results/human_artists/{TEST_IMAGE_NAME}/artist_{ID}.png
.