VISTA-Map/test_gain.py at main · StanfordMSL/VISTA-Map · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
import torch
import open3d as o3d
import numpy as np
import time
import matplotlib
import matplotlib.pyplot as plt
import matplotlib.animation as animation
from pathlib import Path

from vista_map.vista_utils import VistaCoverage
from voxel_traversal.traversal_utils import VoxelGrid

"""
This script visualizes the geometric information gain metric on an example of the Stanford bunny.
"""

def look_at(location, target, up):
    z = (location - target)
    z_ = z / torch.norm(z)
    x = torch.cross(up, z, dim=-1)
    x_ = x / torch.norm(x)
    y = torch.cross(z, x, dim=-1)
    y_ = y / torch.norm(y)

    R = torch.stack([x_, y_, z_], dim=-1)
    return R

# Torch stuff
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.manual_seed(0)

# Load mesh file
obj_path = './vista_map/data/stanford-bunny.obj'

mesh = o3d.io.read_triangle_mesh(obj_path)
mesh.compute_vertex_normals()
mesh.rotate(mesh.get_rotation_matrix_from_xyz((np.pi/2, 0, 0)), center=np.zeros(3))
mesh.translate(-mesh.get_center())
points = np.asarray(mesh.vertices)

# Define params
bounding_box = mesh.get_axis_aligned_bounding_box()
# cmap = matplotlib.cm.get_cmap('turbo')'
cmap = matplotlib.colormaps['turbo']

K = torch.tensor([
    [100., 0., 100.],
    [0., 100., 100.],
    [0., 0., 1.]
], device=device)
far_clip = 1.

param_dict = {
    'discretizations': torch.tensor([100, 100, 100], device=device),
    'lower_bound': torch.tensor(bounding_box.get_min_bound(), device=device),
    'upper_bound': torch.tensor(bounding_box.get_max_bound(), device=device),
    'voxel_grid_values': None,
    'voxel_grid_binary': None,
    'K': K,
    'far_clip': far_clip,
    'val_ndim': 4,
}

normalized_z = (points[:, 2] - points[:, 2].min()) / (points[:, 2].max() - points[:, 2].min())
colors = cmap( normalized_z )[:, :3]
values = np.concatenate([colors, normalized_z.reshape(-1, 1)], axis=-1)

# Initialize voxel grid
vgrid = VoxelGrid(param_dict, 3, device)
vgrid.populate_voxel_grid_from_points(torch.tensor(points, device=device, dtype=torch.float16), torch.tensor(values, device=device, dtype=torch.float16))

# Define "training view" cameras
N_cameras = 100
t = torch.linspace(0., np.pi / 4, N_cameras, device=device)

depth_images = []
directions = []
origins = []

c2w = torch.eye(4, device=device)[None].expand(N_cameras, -1, -1).clone()
c2w[:, :3, 3] = torch.stack([0.25*torch.cos(t), 0.25*torch.sin(t), torch.zeros_like(t)], dim=-1)

target = torch.zeros(3, device=device)[None].expand(N_cameras, -1)
up = torch.tensor([0., 0., 1.], device=device)[None].expand(N_cameras, -1)
c2w[:, :3, :3] = look_at(c2w[:, :3, 3], target, up)

tnow = time.time()
torch.cuda.synchronize()
image, depth, output = vgrid.camera_voxel_intersection(K, c2w, far_clip)
torch.cuda.synchronize()
print("Time taken: ", time.time() - tnow)

# Segment out the rays that intersected the voxel grid
values = image.reshape(-1, image.shape[-1])
depths = depth.reshape(-1)
directions = output['rays_d'].reshape(-1, 3)
origins = output['rays_o'].reshape(-1, 3)

# For depths greater than 0
mask = depths == 0
depths[mask] = far_clip

# Termination points
directions = directions / torch.linalg.norm(directions, axis=-1, keepdims=True)
xyzs = origins + depths[:, None] * directions

pointcloud = {
    'points': xyzs,
    'values': values,
    'origins': origins
}


vista = VistaCoverage(param_dict, pointcloud, 3, device)

# Define test time viewing cameras
N_cameras = 100
t = torch.linspace(0., 2*np.pi, N_cameras, device=device)

c2w = torch.eye(4, device=device)[None].expand(N_cameras, -1, -1).clone()
c2w[:, :3, 3] = torch.stack([0.25*torch.cos(t), 0.25*torch.sin(t), -0.1 + 0.2*torch.arange(len(t), device=device)/len(t)], dim=-1)

target = torch.zeros(3, device=device)[None].expand(N_cameras, -1)
up = torch.tensor([0., 0., 1.], device=device)[None].expand(N_cameras, -1)
c2w[:, :3, :3] = look_at(c2w[:, :3, 3], target, up)

tnow = time.time()
torch.cuda.synchronize()
image, depth, _ = vista.camera_voxel_intersection(K, c2w, far_clip)
torch.cuda.synchronize()
print("Time taken: ", time.time() - tnow)

# Concatenate information gain image with depth image
combined_images_list = []
cmap = matplotlib.colormaps['turbo']

# cmap = matplotlib.cm.get_cmap('turbo')
for i, (img, dep) in enumerate(zip(image, depth)):
    dep = dep / far_clip
    img = img[..., -1]
    combined_image = torch.concatenate([img, dep], axis=1)

    mask = (combined_image == 0).cpu().numpy()
    combined_image = cmap(combined_image.cpu().numpy())[..., :3]
    combined_image[mask] = 0

    combined_images_list.append((combined_image * 255).astype(np.uint8))

# Generate gif of information gain metric
fig_dir = Path("vista_map/images/")

num_images = len(combined_images_list)
fig_comb, axs_comb = plt.subplots()

im_comb = axs_comb.imshow(combined_images_list[0], animated=True)
def animate(i):
    im_comb.set_array(combined_images_list[i])
    return im_comb,

ani_comb = animation.FuncAnimation(fig_comb, animate, frames = num_images, interval=200, blit=True, repeat_delay=10)
ani_comb.save(f"{fig_dir}/vista_coverage.gif")