Single Image Depth Estimation Trained via Depth from Defocus Cues

Official implementation of "Single Image Depth Estimation Trained via Depth from Defocus Cues" (arxiv). The implementation is based on the architectures of DeepLabV3+ and Self-Attention.

Prerequisites

• Python 3.6
• Pytorch 0.4
• Numpy
• Scipy
• OpenCV
• Path
• tqdm
• h5py

Gaussian PSF Layer

In order to build the Gaussian PSF layer on your own machine, run the following line

python ext/setup.py install

Datasets

KITTI

Download the official dataset and use SfmLearner for data preparation.

Make3D

Download the official dataset and unpack into Test and Train folders.

NYU-V2

Download the official dataset and use the following preprocessing instructions.

Flowers

Download the official dataset from here and generate all bokeh images accorfing to Srinivasan.

DSLR

Download the official dataset from here.

Training

CUDA_VISIBLE_DEVICES=0 python train.py --nof-focus 2