Image Depth Masking

An experiment to generate masks for images based off of estimated MiDaS depth

Table of Contents

• Image Depth Masking
  • Table of Contents
  • About
    • Note: Repository Undergoing Maintence
    • Supported Datasets
  • How to Install
    • Packaged Code
    • From Source
  • How to Run
    • COCO specific entrypoints
  • How to Develop/ Extend

About

This is a cam2 project.

It is a usage of the MiDaS depth estimation models to generate masks for images in order to reduce the computational intensity of computer vision (CV) tasks.

Note: Repository Undergoing Maintence

This repository was inherited from Emmanual Amobi ().

I'm currently undergoing a refactoring effort to package and document his work. Please parden the mess.

Supported Datasets

Currently, the following datasets are supported:

• COCO 2014
• COCO 2017

The following datasets are planned to be incorporated:

• MOT
• CIFAR 10
• CIFAR 100
• Google Open Images Dataset

How to Install

Packaged Code

I release Python 3.10.4+ packages of this project here on GitHub.

Get the latest version here and install using pip.

From Source

This project uses poetry as its build tool. You will need it to package this project.

1. git clone https://github.com/NicholasSynovic/image-depth-masking.git
2. cd image-depth-masking
3. poetry update
4. poetry build
5. poetry install dist/*.whl or poetry install dist/*.tar.gz

How to Run

To download datasets for testing, run the specific download{}.bash script where {} is the dataset name. You will need parallel installed to use these scripts.

WARNING: This will download the entire dataset of your choosing which could take up several gigabytes.

To manually download datasets, URLs are provided in the {}URLS.txt files where {} is the dataset name.

Depending on the dataset that you are using, there are different entrypoints to this tool.

COCO specific entrypoints

• idm

  • usage: Find masks for images from the COCO Dataset,
    
    options:
    -a COCO_ANNOTATIONS_FILE, --coco-annotations-file COCO_ANNOTATIONS_FILE
                            A COCO annotations file in JSON format
    -d DEPTH_LEVEL, --depth-level DEPTH_LEVEL
                            The starting depth level to mask at. This should be between 0 and 1. This value is decremented by the
                            depth level decline value until the threshold is met. DEFAULT: 0.9
    -h, --help            show this help message and exit
    -i COCO_IMAGE_FOLDER, --coco-image-folder COCO_IMAGE_FOLDER
                            A path pointing to a folder of images from either the 2014 or 2017 COCO dataset.
    -l DEPTH_LEVEL_DECLINE, --depth-level-decline DEPTH_LEVEL_DECLINE
                            Set value that reduces the depth-level should a mask not be found at that level. This should be between
                            0 and 1. DEFAULT: 0.1
    -m MODEL, --model MODEL
                            A MiDaS compatible model. Supported arguements are: 'DPT_Large', 'DPT_Hybrid', 'MiDaS_small'. DEFAULT:
                            'MiDaS_small'.
    -o OUTPUT_DIRECTORY, --output-directory OUTPUT_DIRECTORY
                            A directory to store masked images. DEFAULT: ./output.
    -s STEPPER, --stepper STEPPER
                            A stepper to step through the image folder. Helps reduce the number of images to be analyzed. DEFAULT:
                            1
    -t THRESHOLD, --threshold THRESHOLD
                            Threshold that must be met for the mask to be valid. This should be between 0 and 1. DEFAULT: 0.9
    -v, --version         Print version of the tool
    
    Tool created by Nicholas M. Synovic <[email protected],Emmanual Amobi <[email protected]>.
    

How to Develop/ Extend

1. git clone https://github.com/NicholasSynovic/image-depth-masking.git
2. cd image-depth-masking
3. poetry update

Running these steps will ensure that you have all of the dependencies installed and configured within a virtual environment that poetry created.

The clean.bash script is useful for cleaning your code and generating reports.