Photography Style Analysis Using Machine Learning

DoF dataset

A novel dataset to quantify image aesthetics and camera bokeh released in various image formats.

• Horizontal (900x600) images: ⬇️download [2.21GB]

• Square (400x400) images: ⬇️download [1.29GB]

• Annotations: ⬇️download

Pretrained models

Pretrained models trained on DoF dataset, in passive and active learning with DenseNet (lite) and pretrained VGG architectures

• Pretrained Models: ⬇️download

Active Learning strategies

Active Learning w/ Incremental Training Active Learning Loop

Active Learning w/ Single Query Active Learning

Repository organization

Install dependency libraries

# Install poetry
$ pip install --user poetry

# Install project's dependencies
$ poetry install

# Enable environment
$ poetry shell

Image transformation

(image_analysis module)$ python im_analysis.py

Training

To initiate training, the following configurations should be available

1. Edit configuration.json

"training_mode": "sqal/sqal_ceal/all/all_ceal"
"ceal": true/false
"model_arch": "densenet/vgg"
"optimizer": "adam/nadam/rms/adagrad/sgd/adadelta"

2. Train model

$ python train_model.py

Note: A CUDA capable GPU with cuda/cuDNN drivers enabled and >=4GB RAM is recommended. TF 2.4 is compatible w/ Cuda 11.0 and cudnn v8.0.5.

NVIDIA's archive: https://developer.nvidia.com/rdp/cudnn-archive