Record3D

Parse and stream Record3D captures. To get demo data, see ../assets/download_assets.sh.

Features:

• viser.extras Record3D parser for RGBD data

• Point cloud visualization from depth maps

• Camera pose trajectory display

• Temporal playback controls with scrubbing

Note

This example requires external assets. To download them, run:

git clone -b v1.0.26 https://github.com/viser-project/viser.git
cd viser/examples
./assets/download_assets.sh
python 04_demos/00_record3d_visualizer.py  # With viser installed.

Source: examples/04_demos/00_record3d_visualizer.py

Code

import time
from pathlib import Path

import numpy as np
import tyro
from tqdm.auto import tqdm

import viser
import viser.extras
import viser.transforms as tf


def main(
    data_path: Path = Path(__file__).parent / "../assets/record3d_dance",
    downsample_factor: int = 4,
    max_frames: int = 100,
    share: bool = False,
) -> None:
    server = viser.ViserServer()
    server.initial_camera.position = (1.2, 1.2, 1.2)
    if share:
        server.request_share_url()

    print("Loading frames!")
    loader = viser.extras.Record3dLoader(data_path)
    num_frames = min(max_frames, loader.num_frames())

    # Initial camera pose.
    @server.on_client_connect
    def _(client: viser.ClientHandle) -> None:
        client.camera.position = (-1.554, -1.013, 1.142)
        client.camera.look_at = (-0.005, 2.283, -0.156)

    # Add playback UI.
    with server.gui.add_folder("Playback"):
        gui_point_size = server.gui.add_slider(
            "Point size",
            min=0.001,
            max=0.02,
            step=1e-3,
            initial_value=0.01,
        )
        gui_timestep = server.gui.add_slider(
            "Timestep",
            min=0,
            max=num_frames - 1,
            step=1,
            initial_value=0,
            disabled=True,
        )
        gui_next_frame = server.gui.add_button("Next Frame", disabled=True)
        gui_prev_frame = server.gui.add_button("Prev Frame", disabled=True)
        gui_playing = server.gui.add_checkbox("Playing", True)
        gui_framerate = server.gui.add_slider(
            "FPS", min=1, max=60, step=0.1, initial_value=loader.fps
        )
        gui_framerate_options = server.gui.add_button_group(
            "FPS options", ("10", "20", "30", "60")
        )

    # Frame step buttons.
    @gui_next_frame.on_click
    def _(_) -> None:
        gui_timestep.value = (gui_timestep.value + 1) % num_frames

    @gui_prev_frame.on_click
    def _(_) -> None:
        gui_timestep.value = (gui_timestep.value - 1) % num_frames

    # Disable frame controls when we're playing.
    @gui_playing.on_update
    def _(_) -> None:
        gui_timestep.disabled = gui_playing.value
        gui_next_frame.disabled = gui_playing.value
        gui_prev_frame.disabled = gui_playing.value

    # Set the framerate when we click one of the options.
    @gui_framerate_options.on_click
    def _(_) -> None:
        gui_framerate.value = int(gui_framerate_options.value)

    prev_timestep = gui_timestep.value

    # Toggle frame visibility when the timestep slider changes.
    @gui_timestep.on_update
    def _(_) -> None:
        nonlocal prev_timestep
        current_timestep = gui_timestep.value
        with server.atomic():
            # Toggle visibility.
            frame_nodes[current_timestep].visible = True
            frame_nodes[prev_timestep].visible = False
        prev_timestep = current_timestep
        server.flush()  # Optional!

    # Load in frames.
    server.scene.add_frame(
        "/frames",
        wxyz=tf.SO3.exp(np.array([np.pi / 2.0, 0.0, 0.0])).wxyz,
        position=(0, 0, 0),
        show_axes=False,
    )
    frame_nodes: list[viser.FrameHandle] = []
    point_nodes: list[viser.PointCloudHandle] = []
    for i in tqdm(range(num_frames)):
        frame = loader.get_frame(i)
        position, color = frame.get_point_cloud(downsample_factor)

        # Add base frame.
        frame_nodes.append(server.scene.add_frame(f"/frames/t{i}", show_axes=False))

        # Place the point cloud in the frame.
        point_nodes.append(
            server.scene.add_point_cloud(
                name=f"/frames/t{i}/point_cloud",
                points=position,
                colors=color,
                point_size=gui_point_size.value,
                point_shape="rounded",
            )
        )

        # Place the frustum.
        fov = 2 * np.arctan2(frame.rgb.shape[0] / 2, frame.K[0, 0])
        aspect = frame.rgb.shape[1] / frame.rgb.shape[0]
        server.scene.add_camera_frustum(
            f"/frames/t{i}/frustum",
            fov=fov,
            aspect=aspect,
            scale=0.15,
            image=frame.rgb[::downsample_factor, ::downsample_factor],
            wxyz=tf.SO3.from_matrix(frame.T_world_camera[:3, :3]).wxyz,
            position=frame.T_world_camera[:3, 3],
        )

        # Add some axes.
        server.scene.add_frame(
            f"/frames/t{i}/frustum/axes",
            axes_length=0.05,
            axes_radius=0.005,
        )

    # Hide all but the current frame.
    for i, frame_node in enumerate(frame_nodes):
        frame_node.visible = i == gui_timestep.value

    # Playback update loop.
    prev_timestep = gui_timestep.value
    while True:
        # Update the timestep if we're playing.
        if gui_playing.value:
            gui_timestep.value = (gui_timestep.value + 1) % num_frames

        # Update point size of both this timestep and the next one! There's
        # redundancy here, but this will be optimized out internally by viser.
        #
        # We update the point size for the next timestep so that it will be
        # immediately available when we toggle the visibility.
        point_nodes[gui_timestep.value].point_size = gui_point_size.value
        point_nodes[
            (gui_timestep.value + 1) % num_frames
        ].point_size = gui_point_size.value

        time.sleep(1.0 / gui_framerate.value)


if __name__ == "__main__":
    tyro.cli(main)