Algorithm data for the 2017 aesthetic value project (NESP 3.2.3, Griffith Institute for Tourism Research)

Brief description

This dataset contains the caffe deep-learning framework along with the setup for image aesthetic train and test code for the Algorithm data. We used NVIDIA-digit 6 environment and this version use caffe 0.15.14 More details information can be found in http://caffe.berkeleyvision.org.

This dataset consists of two folders which are related to automatic aesthetic rating of images using Deep Learning. .
GBR-Aesthetics-Data: This folder contains few sub folders.
1. data-images: - 2500 images used to survey the score.
- 2500 images after resize to 224x224 pixels
- imagenet mean file - a file essential for neural network training
- train list files - list of file names used for training
- test list files – list of file names used for testing
2. lmdb : Two sets of converted images and score into lmdb format. Lmdb format is required during train and test process.
3. tar.gz : tar file contains model definition, trained models and information related to configuration (solver) parameters
4. Qualtrics.xls file contains files names along with their surveyed scores.
5. Infer Many Images.html – Contains generated score from 500 test images using our deep learning model.

GBR-Aesthetics-code: It contains a caffe deep learning framework code.

Methods:
The following step were used to prepare the dataset:
1. Flickr API was used to download more than 10,000 images using a keyword “Great Barrier Reef”. About 10,000 pictures were downloaded in August and September 2017.

2. 2500 images were manually selected to conduct an online survey for manual score assessment based on several research criteria: (i) underwater pictures of GBR, (ii) without humans, (iii) viewed from 1-2 metres from objects and (iv) of high resolution.

3. The survey was created on Qualtrics website and launched on 4th October 2017 using Qualtrics survey service. 772 survey completions were recorded and 705 questionnaires were eligible for data analysis after filtering unqualified questionnaires.

4. At least 10 participants were used to score one picture in a range of 1 to 10. An average score was considered as an actual score.

5. The GBR-aesthetic-code folder actually contains the caffe deep-learning framework along with the setup for image aesthetic train and test code. More details information can be found in http://caffe.berkeleyvision.org. We used NVIDIA-digit 6 environment and this version use caffe 0.15.14


Format:
1. All image files are stored as JPEG (.jpg format) – This images are used for training and testing. However, files are converted to lmdb format before used for actual training process.
2. All deep learning configuration files are saved as recommended prototxt files.
3. lmdb format is used to prepare the final training sets.
4. training and test file lists are stored in txt files.
5. surveyed information are stored in xls files.
6. train_val.prototxt file describes the network definition used for training.
7. solver.protxt contains information related to network configuration parameters
8. snapshot_iter_3360.caffemodel- It is a trained model after 3360 iterations
9. deploy.prototxt- contains network definition of test process.



Data Location:

This dataset is filed in the eAtlas enduring data repository at: data\nesp3\3.2.3_Aesthetic-value-GBR