Depth Estimation from uncalibrated images

About

This is a CLI tool to estimate depth of an anchor image given two pose images. These images can be uncalibrated and unstructured in charateristic. A pipeline of already existing tools has been used in this tool to estimate the depth from three unstructured images. The overall architecture of the tool is shown above.

Getting Started

Prerequisites

You need Python 3.x and Conda package manager to run this tool

Installation

For installing this tool with pretrained model, follow the steps below :

1. Clone this repository git clone https://github.com/tre3x/Depth-Estimation.git
2. Install the neccessary dependencies by executing conda env create -f environment.yml

Usage

To run the tool on local machine, follow the steps in the Installation section. After setting up the environment, Run :

python depthestimate.py --leftimgpath 'path/to/left/image' --anchorimgpath 'path/to/anchor/image' --rightimgpath 'path/to/right/image' --stereo_model 'MSNet2D' --monocular_encoder 'large07' --stereo_ckpt_path 'checkpoint/to/stereo/model' --monocular_checkpoint_path 'checkpoint/to/monocular/depth/estimation/model'

• <path/to/left/image> - Path to pose left image.

• <path/to/anchor/image> - Path to pose anchor image.

• <path/to/right/image> - Path to pose right image.

• <stereo_model> - Type of stereo model to be used in the tool. (MSNet2D/MSNet3D) Default - MSNet2D

• <monocular_encoder> - Type of monocular depth encoder to be used in the tool. (KITTI/NYU) - Default - NYU

• <checkpoint/to/stereo/model> - Path to stereo depth estimation model. Downloadable from here

• <checkpoint/to/monocular/depth/estimation/model> - Path to monocular depth estimation model. Downloadable from here

  To get help about the syntax format : python depthestimate.py --help

Reference

The implementation of the code is based on MobileStereoNet and NewCRFS.