Light Field Distance metric

Original repo: link

Note

The code was converted to be able to use LFD metric (distance between two descriptors) that will compare visual appearance between ground truth mesh and retrieved mesh.

This fork

The original repository was adapted partially to run on Linux. Only LightField was changed so it can be used through docker without any dependency. Underneath, the container uses OSMesa for headless rendering.

Requirements

pip install trimesh

Installation

pip install light-field-distance

python setup.py install

No need to explicitly install anything.

Usage

from lfd import LightFieldDistance
import trimesh

# rest of code
mesh_1: trimesh.Trimesh = ...
mesh_2: trimesh.Trimesh = ...

lfd_value: float = LightFieldDistance(verbose=True).get_distance(
    mesh_1.vertices, mesh_1.faces,
    mesh_2.vertices, mesh_2.faces
)

The script will calculate light field distances [1] between two shapes. Example usage:

from lfd import LightFieldDistance
import trimesh

# rest of code
mesh_1: trimesh.Trimesh = trimesh.load("lfd/examples/cup1.obj")
mesh_2: trimesh.Trimesh = trimesh.load("lfd/examples/airplane.obj")

lfd_value: float = LightFieldDistance(verbose=True).get_distance(
    mesh_1.vertices, mesh_1.faces,
    mesh_2.vertices, mesh_2.faces
)

The lower the metric's value, the more similar shapes are in terms of the visual appearance

How does it work

The lfd.py is a proxy for the container that install all the dependency necessary to run a C code. The code performs calculation of Zernike moments and other coefficients that are necessary to calculate the distance (3DAlignment program). Then, these coefficients are saved and run by the Distance program that calculated the Light Field Distance. It prints out the result and the stdout from the printing is handled by the python script.

If an image for the C code is not found, it builds one. The operation is performed once and it takes a while to finish it. After that, the script runs the necessary computations transparently.

Build with docker

Current repo could be build via Dockerfile. Certain packages assume that render will be done on headless server, but I think it should also work on any PC. Build and run docker via next commands:

docker build -t light-field:latest .
docker run --rm -it -d -v /home:$(pwd) light-field:latest bash

To test and compare cup vs plate:

python3 test.py

For headless render command will be:

Xvfb :99 -screen 0 1900x1080x24+32 & export DISPLAY=:99 && python3 test.py

Contribution

For anyone interested in having a contribution, these are things to be done. Due to the time constraints, I'm not able to do these on my own:

retrieve calculating coefficients from renders to be returned by a method
bind C code with pybind11 to allow direct computation from the python code without any Docker dependency

How am I sure that it works as supposed?

I checked descriptor artifacts from the original implementation and compared with results in the docker through md5sum

TaplierShiru/light-field-distance