Skip to content

Simple VR180 image (format) converter (e.g. equidistant → equirectangular, rotation calibration)

License

Notifications You must be signed in to change notification settings

34j/vr180-convert

Repository files navigation

VR180 image converter

CI Status Documentation Status Test coverage percentage

Poetry black pre-commit

PyPI Version Supported Python versions License


Documentation: https://vr180-convert.readthedocs.io

Source Code: https://github.com/34j/vr180-convert


Simple VR180 image converter on top of OpenCV and NumPy.

Installation

Install this via pip (or your favourite package manager):

pipx install vr180-convert

Usage

Simply run the following command to convert 2 fisheye images to a SBS equirectangular VR180 image:

v1c lr left.jpg right.jpg
left.jpg right.jpg Output
left right output

If left and right image paths are the same, the image is divided into two halves (left and right, SBS) and processed as if they were separate images.

Advanced usage

Automatic image search

If one of left or right image path is a directory, the program will search for the closest image (in terms of creation time) in the other directory.

v1c lr left.jpg right_dir
v1c lr left_dir right.jpg

Since clocks on cameras may not be very accurate in some cases, it is recommended to check how quickly the clocks of the two cameras shift, and synchronize the clocks before shooting. However, it can be adjusted by specifying -ac option.

v1c lr left.jpg right_dir -ac 1 # the clock of the right camera is 1 second faster / ahead
v1c lr left_dir right.jpg -ac 1 # the clock of the right camera is 1 second faster / ahead

Radius estimation

The radius of the non-black area of the input image is assumed by counting black pixels by default, but it would be better to specify it manually to get stable results:

v1c lr left.jpg right.jpg --radius 1000
v1c lr left.jpg right.jpg --radius max # min(width, height) / 2

Calibration

Rotation matching using the least-squares method can be performed by clicking corresponding points that can be regarded as infinitely far away from the camera.

  • Using AKAZE as a feature matcher:
v1c lr left.jpg right.jpg --automatch fm

Since matching by AKAZE involves false detections, matching points with high loss are considered outliers, and the least-squares method is repeated multiple times to remove them. Checking if the image should be swapped using feature matching might be theoretically possible but not implemented.

See Also

v1c lr left.jpg right.jpg --automatch gui
  • Manually specifying the corresponding points using the CLI:
v1c lr left.jpg right.jpg --automatch "0,0;0,0;1,1;1,1" # left_x1,left_y1;right_x1,right_y1;...

$$ a_k, b_k \in \mathbb{R}^3, \min_{R \in SO(3)} \sum_k |R a_k - b_k|^2 $$

Please also refer to the Documentation for mathematical details.

Anaglyph

--merge option (which exports as anaglyph image) can be used to check if the calibration is successful by checking if the infinitely far points are overlapped.

v1c lr left.jpg right.jpg --automatch gui --merge

Swap

If the camera is mounted upside down, you can simply use the --swap option without changing the transformer or other parameters:

v1c lr left.jpg right.jpg --swap

Or the image can be simply swapped using the swap command:

v1c swap rl.jpg

in case one notices that the left and right images are swapped after the conversion.

Convert to Google's format (Photo Sphere XMP Metadata)

This format is special in that it base64-encodes the right-eye image into the metadata of the left-eye image. Required for Google Photos, etc.

You can convert the image to this format by:

v1c xmp lr.jpg

The python-xmp-toolkit used in this command requires exempi to be installed. Note that if this command is called on Windows, it will attempt to install this library and its dependencies and then run the command on WSL using subprocess.

References

Custom conversion model

You can also specify the conversion model by adding Python code directly to the --transformer option:

v1c lr left.jpg right.jpg --transformer 'EquirectangularEncoder() * Euclidean3DRotator(from_rotation_vector([0, np.pi / 4, 0])) * FisheyeDecoder("equidistant")'

If tuple, the first transformer is applied to the left image and the second transformer is applied to the right image. If a single transformer is given, it is applied to both images.

Please refer to the API documentation for the available transformers and their parameters. For from_rotation_vector, please refer to the numpy-quaternion documentation.

Single image conversion

To convert a single image, use v1c s instead.

Running commands for all images in a directory

find left_dir -type f -name '*.jpg' -exec v1c lr {} right_dir --automatch fm --radius max -ac 0 --out-path out \;

Help

For more information, please refer to the help or API documentation:

v1c --help

Usage as a library

For more complex transformations, it is recommended to create your own Transformer.

Note that the transformation is applied in inverse order (new[(x, y)] = old[transform(x, y)], e.g. to decode orthographic fisheye images, transform_polar should be arcsin(theta), not sin(theta).)

from vr180_convert import PolarRollTransformer, apply_lr

class MyTransformer(PolarRollTransformer):
    def transform_polar(
        self, theta: NDArray, roll: NDArray, **kwargs: Any
    ) -> tuple[NDArray, NDArray]:
        return theta**0.98 + theta**1.01, roll

transformer = EquirectangularEncoder() * MyTransformer() * FisheyeDecoder("equidistant")
apply_lr(transformer, left_path="left.jpg", right_path="right.jpg", out_path="output.jpg")

Tips

How to determine which image is left or right

Left Right
Subject Orientation Right Left
Film Color ${\color{red}\text{Red}}$ ${\color{blue}\text{Blue}}$
Anaglyph Color ${\color{blue}\text{Blue}}$ ${\color{red}\text{Red}}$
  • In a SBS image, the subject is oriented toward the center.

How to edit images

This program cannot read RAW files. To deal with white-outs, etc., it is required to process each image with a program such as Photoshop, Lightroom, RawTherapee, Darktable, etc.

However, this is so exhaustive, so it is recommended to take the images with JPEG format with care to avoid overexposure and to match the settings of the two cameras, then convert them with this program and edit the converted images.

Example of editing in RawTherapee (Light editing)
  1. Rank the left images in RawTherapee.
  2. Use this program to convert the images.
  3. Edit the converted images in RawTherapee.
Example of editing in Photoshop (Exquisite editing)
  1. Rank the left images in RawTherapee or Lightroom.
  2. Open left image as a Smart Object LRaw.
  3. Add right image as a Smart Object RRaw.
  4. Make minimal corrections just to match the exposure using Camera Raw Filter.
  5. Make each Smart Object into Smart Objects (L, R) again and do any image-dependent processing, such as removing the background.
  6. Make both images into a single Smart Object (P) and process them as a whole.
  7. Export as a PNG file.
  8. Hide the other Smart Object (L or R) (created in step 3) in the Smart Object P (created in step 4) and save the Smart Object P, then export as a PNG file.
  9. Use this program to convert the images.

Contributors ✨

Thanks goes to these wonderful people (emoji key):

This project follows the all-contributors specification. Contributions of any kind welcome!

Credits

This package was created with Copier and the browniebroke/pypackage-template project template.