Aesthetic Critiques Generation for Photos
Created by Kuang-Yu Chang*, Kung-Hung Lu* and Chu-Song Chen at Academia Sinica, Taipei, Taiwan
This work proposed a paragraph-to-sentence captioning model to generate aesthetic-oriented captions for images. There are various ways to comment an image(e.g. lighting, composition and subject...), especially in aesthetic quality analysis. Owing to the nature of multi-aspect for photo critique, we assume there is an input dataset of images with comments on various aspects. Our approach, aspect fusion(AF) could fuse serveral comments from different aspects for an image and exploit an attention mechenism to generate appropriate photo critique. Moreover, AF could produce more diverse captions than simple CNN-LSTM model and thus is favorable to human. All the training is of end-to-end manner. The implementation is used with Torch framework and based on [neuraltalk2](http://github.com/karpathy/neuraltalk2.
If you find Ordered Weighted Averagin Layer useful in your research, please consider citing:
@inproceedings{chang2017aesthetic,
title={Aesthetic critiques generation for photos},
author={Chang, Kuang-Yu and Lu, Kung-Hung and Chen, Chu-Song},
booktitle={Proceedings of the IEEE International Conference on Computer Vision},
pages={3514--3523},
year={2017}
}
Follow the Neuraltalk2 requirements to set up Torch
run the script train.sh. Use aspect_net to load the pretrained model for the attribute nets and dec_net for the fusion net.
$ ./train.sh
run the script eval.sh. You can set the testing image folder and the number of images you want to eval. If you want to visualize your images in order, you can use an extra index_json file (refer to the sample file vis/vis_label_list.json) to load images in that order. Finally, it will create an vis.json inside the vis folder for visualization.
$ ./eval.sh
-
Download link of PCCD dataset. Google drive
-
Description of PCCD dataset
PCCD ├─────── raw.json # aesthetic comments and the other metadata. ├─────── images # images whose file name corresponds to the "title" field in raw.json. └─────── data_rearrange.py # Users can cherry-pick the metadata they're interested in to reformat the data.
Please feel free to leave suggestions or comments to Kung-Hung Lu ([email protected]), Kuang-Yu Chang ([email protected]) and Chu-Song Chen ([email protected])