unsupervised-depth-estimation

Unsupervised single-shot depth estimation with perceptual reconstruction loss

Implementation of a framework for fully unsupervised single-view depth estimation as proposed in:

Preprint version: #####

Installation

#from github
git clone https://github.com/anger-man/unsupervised-depth-estimation
cd unsupervised-depth-estimation
conda env create --name tf_2.2.0 --file=environment.yml
conda activate tf_2.2.0

Usage

Model architectures and training parameters can be set in config_file.ini. Then:

python train.py --direc path_to_project_folder

Data preparation

RGB images should be given in .jpg-format with dimension dxdx3, where the possible values for spatial resolution d should be powers of 2 (at least 128). Depth images should be given in .npy-format with dimension dxdx1. The evaluation folder should contain some paired samples, where an RGB image is linked by an unique index with its depth counterpart.

./path_to_project_folder/
|--input
|----random_filename1.jpg
|----random_filename2.jpg
|--target
|----random_filename1.npy
|----random_filename2.npy
|--evaluation
|----image_index1.jpg
|----depth_index1.npy
|----image_index2.jpg
|----depth_index2.npy

The directory face_depth gives an example of the needed structure of the project folder. It contains pre-processed samples taken from the Texas 3D Face Recognition database (https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=5483908).