Awesome

fisheye-distortion

Method

This script applies a fisheye distortion to rendered images. The output images can be cropped to a rectangle and resized. The fisheye distortion is the same as detailed in OpenCV 4.4 docs. As mentioned in the docs, the distortion is characterized by 4 parameters, [k1, k2, k3, k4], which are passed via the config file or command line arguments.

Usage

The arguments for the script are passed via the config file. All the parameters in the config file can be overriden from the command line. You must specify:

The input dir in which to search for files,
The filename extention of the input files (eg: .rgb.png, .segments.png)
Whether to use linear interpolation during distortion. For RGB images, using linear interpolation will give smoother results. Other types of images (masks, normals, depth) should never use linear interpolation.

Examples

Distorting RGB images: Use linear interpolation to get smoother results

python apply_fisheye_distortion.py dir.input=samples/ file_ext.input=.rgb.png linear_interpolation=True

Distorting Masks: Use nearest neighbor interpolation to preserve correct values

python apply_fisheye_distortion.py dir.input=samples/ file_ext.input=.segments.png

Saving output files to a different directory. A new directory will be created if it does not exist.

python apply_fisheye_distortion.py dir.output=samples/output

Config file

You can edit the config file, config.yaml, to customize the default parameters:

dir:
  input: samples/
  output: null  # If not specified, outputs files to same dir as input.

file_ext:
  input: .rgb.png  # The files you want to apply distortion to
  info: .info.json  # Contains camera intrinsics

linear_interpolation: False  # Enable Linear for RGB images. Else use nearest-neighbor for masks, normals and depth.
crop_and_resize_output: True  # Disable to get original distorted image. Else output will be of same resolution as input.
workers: 0  # How many processes to use to convert files. 0 for all cores in machine.

distortion_parameters:
  # Ref: https://docs.opencv.org/4.4.0/db/d58/group__calib3d__fisheye.html
  # These are the distortion parameters of a fisheye camera model as defined in the fisheye module of OpenCV 4.4.0
  k1: 0.17149
  k2: -0.27191
  k3: 0.25787
  k4: -0.08054


# Hydra specific params.
hydra:
    output_subdir: null  # Disable saving of config files.
    run:
        dir: .  # Set working dir to current directory

defaults:
    # Disable log files
    - hydra/job_logging: default
    - hydra/hydra_logging: disabled

Install

Pip

This script requires Python 3.7 or greater. The dependencies can be installed via pip:

pip install requirements.txt

Docker

Alternately, you can use the provided dockerfile.

# Build the docker image
bash docker_build.sh

# Run the docker image
bash docker_run.sh

# Inside container
$ python apply_fisheye_distortion.py dir_input=/data

This will process all the images found in the container's /data directory. Modify the docker_run.sh script to change which host directory is mounted to the container's /data.

Detailed Description

From OpenCV docs, here is the description of the fisheye camera model used to create the distortion. The distortion co-efficients [k1, k2, k3, k4] are detailed below: distortion-description