mesh2depth_gpu

Fast depthmap generation using PyOpenGL. Inspired by the CPU-only mesh-to-depth package, this project aims to generate depth data from 3D triangular meshes in Python scripts while leveraging GPU-acceleration. On a Linux system, mesh2depth_gpu supports EGL headless rendering, allowing it to be used without any display device.

Installation

pip install mesh2depth_gpu

Example

import numpy as np
import mesh2depth_gpu as m2d

# camera parameters, type1
# same as https://github.com/daeyun/mesh-to-depth except for the 'is_depth' option.
param1 = {
    "cam_pos": [1, 1, 1],
    "cam_lookat": [0, 0, 0],
    "cam_up": [0, 1, 0],
    "x_fov": 0.349, # End-to-end field of view in radians
    "near": 0.01,
    "far": 100.0,
    "height": 1024,
    "width": 1024,
}

# camera parameters, type2
param2 = {
    "K": np.array([
        750.0, 0.0, 512.0,
        0.0, 750.0, 512.0,
        0.0, 0.0, 1.0
    ]).astype(np.float32), # np.array, 3x3 intrinsic matrix
    "m2c": np.array([
        [1.0, 0.0, 0.0, 0.0],
        [0.0, -1.0, 0.0, 0.0],
        [0.0, 0.0, 1.0, 1.0],
        [0.0, 0.0, 0.0, 1.0]
    ]).astype(np.float32), # np.array, 4x4 extrinsic matrix
    "near": 0.01,
    "far": 100.0,
    "height": 1024,
    "width": 1024,
}

params = [param1, param2]

# load mesh data
vertices = ...  # An array of shape (num_vertices, 3) and type np.float32.
faces = ...  # An array of shape (num_faces, 3) and type np.uint32.

# depthmap generation
depth_maps = m2d.convert(vertices, faces, params, empty_pixel_value=np.nan)

Test

Please follow the instructions in test