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:
cd /path/to/viser/examples/assets
./download_assets.sh
Source: examples/04_demos/01_colmap_visualizer.py

Code¶
1import random
2import time
3from pathlib import Path
4from typing import List
5
6import imageio.v3 as iio
7import numpy as np
8import tyro
9from tqdm.auto import tqdm
10
11import viser
12import viser.transforms as vtf
13from viser.extras.colmap import (
14 read_cameras_binary,
15 read_images_binary,
16 read_points3d_binary,
17)
18
19
20def main(
21 colmap_path: Path = Path(__file__).parent / "../assets/colmap_garden/sparse/0",
22 images_path: Path = Path(__file__).parent / "../assets/colmap_garden/images_8",
23 downsample_factor: int = 2,
24 reorient_scene: bool = True,
25) -> None:
26 server = viser.ViserServer()
27 server.gui.configure_theme(titlebar_content=None, control_layout="collapsible")
28
29 # Load the colmap info.
30 cameras = read_cameras_binary(colmap_path / "cameras.bin")
31 images = read_images_binary(colmap_path / "images.bin")
32 points3d = read_points3d_binary(colmap_path / "points3D.bin")
33
34 points = np.array([points3d[p_id].xyz for p_id in points3d])
35 colors = np.array([points3d[p_id].rgb for p_id in points3d])
36
37 gui_reset_up = server.gui.add_button(
38 "Reset up direction",
39 hint="Set the camera control 'up' direction to the current camera's 'up'.",
40 )
41
42 # Let's rotate the scene so the average camera direction is pointing up.
43 if reorient_scene:
44 average_up = (
45 vtf.SO3(np.array([img.qvec for img in images.values()]))
46 @ np.array([0.0, -1.0, 0.0]) # -y is up in the local frame!
47 ).mean(axis=0)
48 average_up /= np.linalg.norm(average_up)
49 server.scene.set_up_direction((average_up[0], average_up[1], average_up[2]))
50
51 @gui_reset_up.on_click
52 def _(event: viser.GuiEvent) -> None:
53 client = event.client
54 assert client is not None
55 client.camera.up_direction = vtf.SO3(client.camera.wxyz) @ np.array(
56 [0.0, -1.0, 0.0]
57 )
58
59 gui_points = server.gui.add_slider(
60 "Max points",
61 min=1,
62 max=len(points3d),
63 step=1,
64 initial_value=min(len(points3d), 50_000),
65 )
66 gui_frames = server.gui.add_slider(
67 "Max frames",
68 min=1,
69 max=len(images),
70 step=1,
71 initial_value=min(len(images), 50),
72 )
73 gui_point_size = server.gui.add_slider(
74 "Point size", min=0.01, max=0.1, step=0.001, initial_value=0.02
75 )
76
77 point_mask = np.random.choice(points.shape[0], gui_points.value, replace=False)
78 point_cloud = server.scene.add_point_cloud(
79 name="/colmap/pcd",
80 points=points[point_mask],
81 colors=colors[point_mask],
82 point_size=gui_point_size.value,
83 )
84 frames: List[viser.FrameHandle] = []
85
86 def visualize_frames() -> None:
87
88 # Remove existing image frames.
89 for frame in frames:
90 frame.remove()
91 frames.clear()
92
93 # Interpret the images and cameras.
94 img_ids = [im.id for im in images.values()]
95 random.shuffle(img_ids)
96 img_ids = sorted(img_ids[: gui_frames.value])
97
98 for img_id in tqdm(img_ids):
99 img = images[img_id]
100 cam = cameras[img.camera_id]
101
102 # Skip images that don't exist.
103 image_filename = images_path / img.name
104 if not image_filename.exists():
105 continue
106
107 T_world_camera = vtf.SE3.from_rotation_and_translation(
108 vtf.SO3(img.qvec), img.tvec
109 ).inverse()
110 frame = server.scene.add_frame(
111 f"/colmap/frame_{img_id}",
112 wxyz=T_world_camera.rotation().wxyz,
113 position=T_world_camera.translation(),
114 axes_length=0.1,
115 axes_radius=0.005,
116 )
117 frames.append(frame)
118
119 # For pinhole cameras, cam.params will be (fx, fy, cx, cy).
120 if cam.model != "PINHOLE":
121 print(f"Expected pinhole camera, but got {cam.model}")
122
123 H, W = cam.height, cam.width
124 fy = cam.params[1]
125 image = iio.imread(image_filename)
126 image = image[::downsample_factor, ::downsample_factor]
127 frustum = server.scene.add_camera_frustum(
128 f"/colmap/frame_{img_id}/frustum",
129 fov=2 * np.arctan2(H / 2, fy),
130 aspect=W / H,
131 scale=0.15,
132 image=image,
133 )
134
135 @frustum.on_click
136 def _(_, frame=frame) -> None:
137 for client in server.get_clients().values():
138 client.camera.wxyz = frame.wxyz
139 client.camera.position = frame.position
140
141 need_update = True
142
143 @gui_points.on_update
144 def _(_) -> None:
145 point_mask = np.random.choice(points.shape[0], gui_points.value, replace=False)
146 with server.atomic():
147 point_cloud.points = points[point_mask]
148 point_cloud.colors = colors[point_mask]
149
150 @gui_frames.on_update
151 def _(_) -> None:
152 nonlocal need_update
153 need_update = True
154
155 @gui_point_size.on_update
156 def _(_) -> None:
157 point_cloud.point_size = gui_point_size.value
158
159 while True:
160 if need_update:
161 need_update = False
162 visualize_frames()
163
164 time.sleep(1e-3)
165
166
167if __name__ == "__main__":
168 tyro.cli(main)