COLMAP

Visualize COLMAP sparse reconstruction outputs. To get demo data, see ../assets/download_assets.sh.

Features:

• COLMAP sparse reconstruction file parsing

• Camera frustum visualization with viser.SceneApi.add_camera_frustum()

• 3D point cloud display from structure-from-motion

• Interactive camera and point visibility controls

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/01_colmap_visualizer.py  # With viser installed.

Source: examples/04_demos/01_colmap_visualizer.py

Code

import random
import time
from pathlib import Path
from typing import List

import imageio.v3 as iio
import numpy as np
import tyro
from tqdm.auto import tqdm

import viser
import viser.transforms as vtf
from viser.extras.colmap import (
    read_cameras_binary,
    read_images_binary,
    read_points3d_binary,
)


def main(
    colmap_path: Path = Path(__file__).parent / "../assets/colmap_garden/sparse/0",
    images_path: Path = Path(__file__).parent / "../assets/colmap_garden/images_8",
    downsample_factor: int = 2,
    reorient_scene: bool = True,
) -> None:
    server = viser.ViserServer()
    server.gui.configure_theme(titlebar_content=None, control_layout="collapsible")

    # Load the colmap info.
    cameras = read_cameras_binary(colmap_path / "cameras.bin")
    images = read_images_binary(colmap_path / "images.bin")
    points3d = read_points3d_binary(colmap_path / "points3D.bin")

    points = np.array([points3d[p_id].xyz for p_id in points3d])
    colors = np.array([points3d[p_id].rgb for p_id in points3d])

    # Let's rotate the scene so the average camera direction is pointing up.
    if reorient_scene:
        average_up = (
            # `qvec` corresponds to T_camera_world; we convert to T_world_camera.
            vtf.SO3(np.array([img.qvec for img in images.values()])).inverse()
            @ np.array([0.0, -1.0, 0.0])  # -y is up in the local frame!
        ).mean(axis=0)
        average_up /= np.linalg.norm(average_up)
        server.scene.set_up_direction((average_up[0], average_up[1], average_up[2]))

    gui_points = server.gui.add_slider(
        "Max points",
        min=1,
        max=len(points3d),
        step=1,
        initial_value=min(len(points3d), 50_000),
    )
    gui_frames = server.gui.add_slider(
        "Max frames",
        min=1,
        max=len(images),
        step=1,
        initial_value=min(len(images), 50),
    )
    gui_point_size = server.gui.add_slider(
        "Point size", min=0.01, max=0.1, step=0.001, initial_value=0.02
    )

    point_mask = np.random.choice(points.shape[0], gui_points.value, replace=False)
    point_cloud = server.scene.add_point_cloud(
        name="/colmap/pcd",
        points=points[point_mask],
        colors=colors[point_mask],
        point_size=gui_point_size.value,
    )
    frames: List[viser.FrameHandle] = []

    def visualize_frames() -> None:

        # Remove existing image frames.
        for frame in frames:
            frame.remove()
        frames.clear()

        # Interpret the images and cameras.
        img_ids = [im.id for im in images.values()]
        random.shuffle(img_ids)
        img_ids = sorted(img_ids[: gui_frames.value])

        for img_id in tqdm(img_ids):
            img = images[img_id]
            cam = cameras[img.camera_id]

            # Skip images that don't exist.
            image_filename = images_path / img.name
            if not image_filename.exists():
                continue

            T_world_camera = vtf.SE3.from_rotation_and_translation(
                vtf.SO3(img.qvec), img.tvec
            ).inverse()
            frame = server.scene.add_frame(
                f"/colmap/frame_{img_id}",
                wxyz=T_world_camera.rotation().wxyz,
                position=T_world_camera.translation(),
                axes_length=0.1,
                axes_radius=0.005,
            )
            frames.append(frame)

            # For pinhole cameras, cam.params will be (fx, fy, cx, cy).
            if cam.model != "PINHOLE":
                print(f"Expected pinhole camera, but got {cam.model}")

            H, W = cam.height, cam.width
            fy = cam.params[1]
            image = iio.imread(image_filename)
            image = image[::downsample_factor, ::downsample_factor]
            frustum = server.scene.add_camera_frustum(
                f"/colmap/frame_{img_id}/frustum",
                fov=2 * np.arctan2(H / 2, fy),
                aspect=W / H,
                scale=0.15,
                image=image,
            )

            @frustum.on_click
            def _(_, frame=frame) -> None:
                for client in server.get_clients().values():
                    client.camera.wxyz = frame.wxyz
                    client.camera.position = frame.position

    need_update = True

    @gui_points.on_update
    def _(_) -> None:
        point_mask = np.random.choice(points.shape[0], gui_points.value, replace=False)
        with server.atomic():
            point_cloud.points = points[point_mask]
            point_cloud.colors = colors[point_mask]

    @gui_frames.on_update
    def _(_) -> None:
        nonlocal need_update
        need_update = True

    @gui_point_size.on_update
    def _(_) -> None:
        point_cloud.point_size = gui_point_size.value

    while True:
        if need_update:
            need_update = False
            visualize_frames()

        time.sleep(1e-3)


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