-
Notifications
You must be signed in to change notification settings - Fork 0
/
multiview_model.py
192 lines (170 loc) · 8.92 KB
/
multiview_model.py
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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import kornia
#from MCMT.resnet import resnet18
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings("ignore")
class MultiView_Detection(nn.Module):
def __init__(self, backbone_model, logdir, loss, avgpool, cam_set, len_cam_set):
super().__init__()
self.logdir = logdir
self.avgpool = avgpool
self.cam_set = cam_set
self.MAX_CAM = 8
#self.dataset = dataset
#print(dataset.dicts)
'''
self.num_cam = dataset.num_cam
self.img_shape = dataset.img_shape
self.reducedgrid_shape = dataset.reducedgrid_shape
self.upsample_shape = list(map(lambda x: int(x / dataset.img_reduce), self.img_shape))
img_reduce = np.array(self.img_shape) / np.array(self.upsample_shape)
img_zoom_mat = np.diag(np.append(img_reduce, [1]))
map_zoom_mat = np.diag(np.append(np.ones([2]) / dataset.grid_reduce, [1]))
imgcoord2worldgrid_matrices = self.get_imgcoord2worldgrid_matrices(dataset.base.intrinsic_matrices,
dataset.base.extrinsic_matrices,
dataset.base.worldgrid2worldcoord_mat)
# Projection matrix
self.proj_mats = [torch.from_numpy(map_zoom_mat @ imgcoord2worldgrid_matrices[cam] @ img_zoom_mat)
for cam in range(self.num_cam)]
# Coordinate Map
self.coord_map = self.get_coord_map(self.reducedgrid_shape + [1])
'''
# Resnet18
self.base_arch = nn.Sequential(*list(backbone_model.children())[:-2]).to('cuda:0')
'''
for param in base_arch.parameters():
param.requires_grad = False
'''
#split = 7
#self.base_pt1 = base_arch[:split].to('cuda:1')
#self.base_pt2 = base_arch[split:].to('cuda:0')
if avgpool:
self.out_channel = 512+2
else:
if self.cam_set:
self.out_channel = 512 * len_cam_set + 2
else:
self.out_channel = 512 * self.MAX_CAM + 2
# Ground Plane Convolution
if loss == 'klcc':
#### for KLDiv+CC ####
self.map_classifier = nn.Sequential(nn.Conv2d(self.out_channel, 512, 3, padding=1), nn.ReLU(),
nn.Conv2d(512, 512, 3, padding=2, dilation=2), nn.ReLU(),
nn.Conv2d(512, 1, 3, padding=4, dilation=4), nn.Sigmoid()).to('cuda:1')
#############################################
elif loss == 'mse':
self.map_classifier = nn.Sequential(nn.Conv2d(self.out_channel, 512, 3, padding=1), nn.ReLU(),
nn.Conv2d(512, 512, 3, padding=2, dilation=2), nn.ReLU(),
nn.Conv2d(512, 1, 3, padding=4, dilation=4, bias=False)).to('cuda:1')
def forward(self, dataset, dataset_name, imgs, ignore_cam, duplicate_cam, random, cam_selected):
B, N, C, H, W = imgs.shape
#print(dataset.dicts.keys())
#print(dataset_name)
config_dict = dataset.dicts[dataset_name[0]]
#print(config_dict)
num_cam = config_dict['num_cam']
upsample_shape = config_dict['upsample_shape']
reducedgrid_shape = config_dict['reducedgrid_shape']
img_reduce = config_dict['img_reduce']
proj_mats = config_dict['proj_mats']
coord_map = config_dict['coord_map']
assert N == num_cam
#device = imgs.device
world_features = []
#world_features = torch.zeros(B,self.out_channel,self.reducedgrid_shape[0], self.reducedgrid_shape[1]).to(device)
if self.cam_set:
n_cam = len(cam_selected)
else:
n_cam = num_cam
imgs_result = []
for cam in range(n_cam):
if self.cam_set:
cam = cam_selected[cam]
if random:
#print('ignore : ',ignore_cam)
if ignore_cam == cam:
if self.avgpool:
continue
else:
#print('duplicate : ',duplicate_cam)
cam = duplicate_cam
# imgs[batch, cam, channel, h, w]
#print(cam)
img_feature = self.base_arch(imgs[:, cam].to('cuda:0'))
#img_feature = self.base_pt2(img_feature.to('cuda:0'))
img_feature = F.interpolate(img_feature, upsample_shape, mode='bilinear')
'''
fig = plt.figure()
subplt0 = fig.add_subplot(211, title="output")
subplt0.imshow(torch.norm(img_feature[0].detach(), dim=0).cpu().numpy())
plt.savefig(os.path.join(self.logdir,'feature_sc2_map'+str(cam+1)+'.jpg'))
plt.close(fig)
'''
# 3x3 proj mat of cam --> repeat will make it [B,3,3]
proj_mat = proj_mats[cam].repeat([B, 1, 1]).float().to('cuda:1')
if self.avgpool:
world_feature = kornia.warp_perspective(img_feature.to('cuda:1'), proj_mat, reducedgrid_shape).unsqueeze(0)
'''
fig = plt.figure()
subplt0 = fig.add_subplot(211, title="output")
subplt0.imshow(torch.norm(world_feature[0][0].detach(), dim=0).cpu().numpy())
plt.savefig(os.path.join(self.logdir,'world_sc2_map'+str(cam+1)+'.jpg'))
plt.close(fig)
'''
else:
world_feature = kornia.warp_perspective(img_feature.to('cuda:1'), proj_mat, reducedgrid_shape)
######### Concetenate features ###########
world_features.append(world_feature)
##########################################
#world_features += world_feature
# torch.cat(list_of_tensors + list_of_tensor, dim=1)
# torch.cat(world_features + [self.coord_map.repeat([B, 1, 1, 1], dim=1)
# world_features[B, num_camera*channel+2, H, W]
#world_features /= self.num_cam
#world_features = torch.cat([world_features] + [self.coord_map.repeat([B, 1, 1, 1]).to('cuda:0')], dim=1)
#world_features = torch.cat(world_features + [self.coord_map.repeat([B, 1, 1, 1]).to('cuda:0')], dim=1)
if self.avgpool:
world_features = torch.cat(world_features, dim=1)
world_features = torch.mean(world_features, dim=1)
world_features = torch.cat([world_features] + [coord_map.repeat([B, 1, 1, 1]).to('cuda:1')], dim=1)
else:
if num_cam < self.MAX_CAM:
replicate_num = self.MAX_CAM - num_cam
zero_vector = torch.zeros((1,512*replicate_num,reducedgrid_shape[0], reducedgrid_shape[1])).to('cuda:1')
world_features = torch.cat(world_features + [zero_vector] +[coord_map.repeat([B, 1, 1, 1]).to('cuda:1')], dim=1)
else:
world_features = torch.cat(world_features + [coord_map.repeat([B, 1, 1, 1]).to('cuda:1')], dim=1)
fig = plt.figure()
subplt0 = fig.add_subplot(211, title="output")
subplt0.imshow(torch.norm(world_features[0].detach(), dim=0).cpu().numpy())
plt.savefig(os.path.join(self.logdir,'world_features.jpg'))
plt.close(fig)
map_result = self.map_classifier(world_features.to('cuda:1'))
map_result = F.interpolate(map_result, reducedgrid_shape, mode='bilinear')
return map_result
def get_imgcoord2worldgrid_matrices(self, intrinsic_matrices, extrinsic_matrices, worldgrid2worldcoord_mat):
projection_matrices = {}
for cam in range(self.num_cam):
# removing third column(z=0) from extrinsic matrix of size 3x4
worldcoord2imgcoord_mat = intrinsic_matrices[cam] @ np.delete(extrinsic_matrices[cam], 2, 1)
worldgrid2imgcoord_mat = worldcoord2imgcoord_mat @ worldgrid2worldcoord_mat
imgcoord2worldgrid_mat = np.linalg.inv(worldgrid2imgcoord_mat)
# transforming img axis to grid map axis
# x(col),y(row) img coord --> y(col), x(row) grid map coord
permutation_mat = np.array([[0, 1, 0], [1, 0, 0], [0, 0, 1]])
projection_matrices[cam] = permutation_mat @ imgcoord2worldgrid_mat
return projection_matrices
def get_coord_map(self, grid_shape):
H, W, C = grid_shape
# img[y,x] = img[h,w]
grid_x, grid_y = np.meshgrid(np.arange(W), np.arange(H))
# making x and y in range [-1.0 to 1.0]
grid_x = torch.from_numpy((grid_x / (W - 1) * 2) - 1).float()
grid_y = torch.from_numpy((grid_y / (H - 1) * 2) - 1).float()
ret = torch.stack([grid_x, grid_y], dim=0).unsqueeze(0)
return ret