Skip to content

Depth image saved via render_to_depth_image does not match RGB image #6270

Open
@ghweber

Description

@ghweber

Checklist

Describe the issue

I have been noticing that on my Mac (MacBook Pro (14-inch, 2021), macOS version 12.6.7 (21G651)) that the mouse_and_point_coord.py example does not work properly. Points returned when clicking on the image are incorrect. By modifying the example (see below), I was able to trace this to an incorrect depth image (which seems to have an incorrect camera setup) returned by self.widget3d.scene.scene.render_to_depth_image(depth_callback). Calling self.widget3d.scene.scene.render_to_image(image_callback) on the other hand returns a correct image.

Steps to reproduce the bug

# ----------------------------------------------------------------------------
# -                        Open3D: www.open3d.org                            -
# ----------------------------------------------------------------------------
# Copyright (c) 2018-2023 www.open3d.org
# SPDX-License-Identifier: MIT
# ----------------------------------------------------------------------------

import matplotlib.pyplot as plt
import numpy as np
import open3d as o3d
import open3d.visualization.gui as gui
import open3d.visualization.rendering as rendering


# This example displays a point cloud and if you Ctrl-click on a point
# (Cmd-click on macOS) it will show the coordinates of the point.
# This example illustrates:
# - custom mouse handling on SceneWidget
# - getting a the depth value of a point (OpenGL depth)
# - converting from a window point + OpenGL depth to world coordinate
class ExampleApp:

    def __init__(self, cloud):
        # We will create a SceneWidget that fills the entire window, and then
        # a label in the lower left on top of the SceneWidget to display the
        # coordinate.
        app = gui.Application.instance
        self.window = app.create_window("Open3D - GetCoord Example", 1024, 768)
        # Since we want the label on top of the scene, we cannot use a layout,
        # so we need to manually layout the window's children.
        self.window.set_on_layout(self._on_layout)
        self.widget3d = gui.SceneWidget()
        self.window.add_child(self.widget3d)
        self.info = gui.Label("")
        self.info.visible = False
        self.window.add_child(self.info)

        self.widget3d.scene = rendering.Open3DScene(self.window.renderer)

        mat = rendering.MaterialRecord()
        mat.shader = "defaultUnlit"
        # Point size is in native pixels, but "pixel" means different things to
        # different platforms (macOS, in particular), so multiply by Window scale
        # factor.
        mat.point_size = 3 * self.window.scaling
        self.widget3d.scene.add_geometry("Point Cloud", cloud, mat)

        bounds = self.widget3d.scene.bounding_box
        center = bounds.get_center()
        self.widget3d.setup_camera(60, bounds, center)
        self.widget3d.look_at(center, center - [0, 0, 3], [0, -1, 0])

        self.widget3d.set_on_mouse(self._on_mouse_widget3d)

    def _on_layout(self, layout_context):
        r = self.window.content_rect
        self.widget3d.frame = r
        pref = self.info.calc_preferred_size(layout_context,
                                             gui.Widget.Constraints())
        self.info.frame = gui.Rect(r.x,
                                   r.get_bottom() - pref.height, pref.width,
                                   pref.height)

    def _on_mouse_widget3d(self, event):
        # We could override BUTTON_DOWN without a modifier, but that would
        # interfere with manipulating the scene.
        if event.type == gui.MouseEvent.Type.BUTTON_DOWN and event.is_modifier_down(
                gui.KeyModifier.CTRL):

            def depth_callback(depth_image):
                # Coordinates are expressed in absolute coordinates of the
                # window, but to dereference the image correctly we need them
                # relative to the origin of the widget. Note that even if the
                # scene widget is the only thing in the window, if a menubar
                # exists it also takes up space in the window (except on macOS).
                x = event.x - self.widget3d.frame.x
                y = event.y - self.widget3d.frame.y
                depth_array = np.asarray(depth_image)
                print(f"Frame: {self.widget3d.frame.width}x{self.widget3d.frame.height}")
                print(f"Array shape: {depth_array.shape}")
                plt.imsave("depth_buffer.png", depth_array)
                # Note that np.asarray() reverses the axes.
                depth = depth_array[y, x]

                if depth == 1.0:  # clicked on nothing (i.e. the far plane)
                    text = ""
                else:
                    world = self.widget3d.scene.camera.unproject(
                        x, y, depth, self.widget3d.frame.width,
                        self.widget3d.frame.height)
                    text = "({:.3f}, {:.3f}, {:.3f})".format(
                        world[0], world[1], world[2])

                # This is not called on the main thread, so we need to
                # post to the main thread to safely access UI items.
                def update_label():
                    self.info.text = text
                    self.info.visible = (text != "")
                    # We are sizing the info label to be exactly the right size,
                    # so since the text likely changed width, we need to
                    # re-layout to set the new frame.
                    self.window.set_needs_layout()

                gui.Application.instance.post_to_main_thread(
                    self.window, update_label)

            self.widget3d.scene.scene.render_to_depth_image(depth_callback)
            def image_callback(image):
                plt.imsave("image.png", np.asarray(image))
            self.widget3d.scene.scene.render_to_image(image_callback)
            return gui.Widget.EventCallbackResult.HANDLED
        return gui.Widget.EventCallbackResult.IGNORED


def main():
    app = gui.Application.instance
    app.initialize()

    # This example will also work with a triangle mesh, or any 3D object.
    # If you use a triangle mesh you will probably want to set the material
    # shader to "defaultLit" instead of "defaultUnlit".
    pcd_data = o3d.data.DemoICPPointClouds()
    cloud = o3d.io.read_point_cloud(pcd_data.paths[0])
    ex = ExampleApp(cloud)

    app.run()


if __name__ == "__main__":
    main()

Error message

No error message.
Screen shot of window.
Screen Shot 2023-07-17 at 8 09 44 PM
Depth image saved upon mouse click.
depth_buffer
Image saved upon mouse click.
image

Expected behavior

Expect the depth image to align with the color image.

Open3D, Python and System information

- Operating system: macOS 12.6.7 (21G651)
- Python version: 3.9.16 (main, Jul 17 2023, 19:26:31)  [Clang 14.0.0 (clang-1400.0.29.202)
- Open3D version: output from python: 0.17.0
- System architecture: apple-silicon
- Is this a remote workstation?: no
- How did you install Open3D?: pip

Additional information

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugNot a build issue, this is likely a bug.rendering

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions