A neural architecture for scenery image outpaiting (ICCV 2019), implemented in TensorFlow.
The architecture has an ability to generate a very long high-quality prediction from a small input image by outpaiting:
Please install TensorFlow>=1.3.0, Python>=3.6.
For training and testing, we collect a new outpainting dataset, which has 6,000 images containing complex natural scenes. You can download the raw dataset from here and split the training and testing set by yourself. Or, you can get our split from here (TFRecord format, 128 resolution, 5,000 images for training and 1,000 for testing).
For training and evaluation, you can use train.sh and eval.sh. Please remember to set the TFRecord dataset path inside them.
Besides, you can get our pretrain model from here, and run eval_model.py to evaluate it.
After running eval_model.py, the evaluation process will store 4 types of images: 1) "ori_xxx.jpg", the groundtruth images of size 128x256; 2) "m0_xxx.jpg", the 1-step predictions of size 128x256 without any post-processing methods; 3) "m1_xxx.jpg", the 1-step predictions of size 128x256 with smoothly stitching; 4) "endless_xxx.jpg", the 4-step predictions of size 128x640.
Notably, we measure Inception Score and Inception Distance between "ori_xxx.jpg" and "m0_xxx.jpg" in our paper.
@inproceedings{yang2019very,
title={Very Long Natural Scenery Image Prediction by Outpainting},
author={Yang, Zongxin and Dong, Jian and Liu, Ping and Yang, Yi and Yan, Shuicheng},
booktitle={Proceedings of the IEEE International Conference on Computer Vision},
pages={10561--10570},
year={2019}
}