1. Overview  [Paper] [Project Webpage] [PyNET PyTorch]

This repository provides the implementation of the deep learning-based bokeh effect rendering approach presented in this paper. The model is trained to map the standard narrow-aperture images into shallow depth-of-field photos captured with a professional Canon 7D DSLR camera. The presented approach is camera independent, does not require any special hardware, and can also be applied to the existing images. More visual results of this method on the presented EBB! dataset and its comparison to the Portrait Mode of the Google Pixel Camera app can be found here.

2. Prerequisites

3. First steps

  • Download the pre-trained VGG-19 model Mirror and put it into vgg_pretrained/ folder.

  • Download the pre-trained PyNET model and put it into models/original/ folder.

  • Download the EBB! dataset and extract it into ebb_dataset/ folder.
    This folder should contain two subfolders: train/ and test/

    Please note that Google Drive has a quota limiting the number of downloads per day. To avoid it, you can login to your Google account and press "Add to My Drive" button instead of a direct download. Please check this issue for more information.



The proposed PyNET-based architecture has an inverted pyramidal shape and is processing the images at seven different scales (levels). The model is trained sequentially, starting from the lowest 7th layer, which allows to achieve good semantically-driven reconstruction results at smaller scales that are working with images of very low resolution and thus performing mostly global image manipulations. After the bottom layer is pre-trained, the same procedure is applied to the next level till the training is done on the original resolution. Since each higher level is getting upscaled high-quality features from the lower part of the model, it mainly learns to reconstruct the missing low-level details and refines the results. In this work, we additionally use two transposed convolutional layers on top of the main model (Levels 1, 2) that upsample the images to their target size.

5. Training the model

The model is trained level by level, starting from the lowest (7th) one:

python level=<level>

Obligatory parameters:

level: 7, 6, 5, 4, 3, 2, 1

Optional parameters and their default values:

batch_size: 50   -   batch size [small values can lead to unstable training]

train_size: 4894   -   the number of training images randomly loaded each 1000 iterations

eval_step: 1000   -   each eval_step iterations the accuracy is computed and the model is saved

learning_rate: 5e-5   -   learning rate

restore_iter: None   -   iteration to restore (when not specified, the last saved model for PyNET's level+1 is loaded)

num_train_iters: 5K, 5K, 20K, 20K, 30K, 80K, 100K (for levels 5 - 0)   -   the number of training iterations

vgg_dir: vgg_pretrained/imagenet-vgg-verydeep-19.mat   -   path to the pre-trained VGG-19 network

dataset_dir: ebb_dataset/   -   path to the folder with the EBB! dataset

Below we provide the commands used for training the model on the Nvidia Tesla V100 GPU with 16GB of RAM. When using GPUs with smaller amount of memory, the batch size and the number of training iterations should be adjusted accordingly:

python level=7 batch_size=50 num_train_iters=5000
python level=6 batch_size=50 num_train_iters=5000
python level=5 batch_size=40 num_train_iters=20000
python level=4 batch_size=14 num_train_iters=20000
python level=3 batch_size=9 num_train_iters=30000
python level=2 batch_size=9 num_train_iters=80000
python level=1 batch_size=5 num_train_iters=100000

6. Test the provided pre-trained models on full-resolution test EBB! images

python orig=true

Optional parameters:

use_gpu: true,false   -   run the model on GPU or CPU

dataset_dir: ebb_dataset/   -   path to the folder with the EBB! dataset

7. Validate the obtained model on full-resolution test EBB! images


Optional parameters:

restore_iter: None   -   iteration to restore (when not specified, the last saved model for level=`````` is loaded)

use_gpu: true,false   -   run the model on GPU or CPU

dataset_dir: ebb_dataset/   -   path to the folder with the EBB! dataset

8. Folder structure

models/   -   logs and models that are saved during the training process

models/original/   -   the folder with the provided pre-trained PyNET model

ebb_dataset/   -   the folder with the EBB! dataset

results/   -   visual results for image crops that are saved while training

results/full-resolution/   -   full-resolution image results saved during the testing

vgg-pretrained/   -   the folder with the pre-trained VGG-19 network   -   python script that loads training data   -   PyNET implementation (TensorFlow)   -   implementation of the training procedure   -   applying the pre-trained model to full-resolution test images and computing the numerical results   -   auxiliary functions   -   loading the pre-trained vgg-19 network