-
Notifications
You must be signed in to change notification settings - Fork 1
/
demo.py
182 lines (130 loc) · 7.51 KB
/
demo.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
import numpy as np
import torch
import torch.nn as nn
def masked_shift_and_scale(depth_preds, depth_gt, mask_valid):
depth_preds_nan = depth_preds.clone()
depth_gt_nan = depth_gt.clone()
depth_preds_nan[~mask_valid] = np.nan
depth_gt_nan[~mask_valid] = np.nan
mask_diff = mask_valid.reshape(mask_valid.size()[:2] + (-1,)).sum(-1, keepdims=True) + 1
t_gt = depth_gt_nan.reshape(depth_gt_nan.size()[:2] + (-1,)).nanmedian(-1, keepdims=True)[0].unsqueeze(-1)
t_gt[torch.isnan(t_gt)] = 0
diff_gt = torch.abs(depth_gt - t_gt)
diff_gt[~mask_valid] = 0
s_gt = (diff_gt.reshape(diff_gt.size()[:2] + (-1,)).sum(-1, keepdims=True) / mask_diff).unsqueeze(-1)
depth_gt_aligned = (depth_gt - t_gt) / (s_gt + 1e-6)
t_pred = depth_preds_nan.reshape(depth_preds_nan.size()[:2] + (-1,)).nanmedian(-1, keepdims=True)[0].unsqueeze(-1)
t_pred[torch.isnan(t_pred)] = 0
diff_pred = torch.abs(depth_preds - t_pred)
diff_pred[~mask_valid] = 0
s_pred = (diff_pred.reshape(diff_pred.size()[:2] + (-1,)).sum(-1, keepdims=True) / mask_diff).unsqueeze(-1)
depth_pred_aligned = (depth_preds - t_pred) / (s_pred + 1e-6)
return depth_pred_aligned, depth_gt_aligned
def masked_l1_loss(preds, target, mask_valid,dense=False):
element_wise_loss = abs(preds - target)
element_wise_loss[~mask_valid] = 0
if dense is False:
return element_wise_loss.sum() / mask_valid.sum()
else: # not average
return element_wise_loss
def get_contexts_dr( level, depth_gt, mask_valid):
batch_norm_context = []
for mask_index in range(depth_gt.shape[0]): #process each img in the batch
depth_map = depth_gt[mask_index]
valid_map = mask_valid[mask_index]
if depth_map[valid_map].numel() == 0: #if there is no valid pixel
map_context_list = [valid_map for _ in range(2 ** (level) - 1)]
else:
valid_values = depth_map[valid_map]
max_d = valid_values.max()
min_d = valid_values.min()
bin_size_list = [(1 / 2) ** (i) for i in range(level)]
bin_size_list.reverse()
map_context_list = []
for bin_size in bin_size_list:
for i in range(int(1 / bin_size)):
mask_new = (depth_map >= min_d + (max_d - min_d) * i * bin_size) & (
depth_map < min_d + (max_d - min_d) * (i + 1) * bin_size + 1e-30)
mask_new = mask_new & valid_map
map_context_list.append(mask_new)
map_context_list = torch.stack(map_context_list, dim=0)
batch_norm_context.append(map_context_list)
batch_norm_context = torch.stack(batch_norm_context, dim=0).swapdims(0, 1)
return batch_norm_context
def get_contexts_dp( level, depth_gt, mask_valid):
depth_gt_nan=depth_gt.clone()
depth_gt_nan[~mask_valid] = np.nan
depth_gt_nan=depth_gt_nan.view(depth_gt_nan.shape[0], depth_gt_nan.shape[1], -1)
bin_size_list = [(1 / 2) ** (i) for i in range(level)]
bin_size_list.reverse()
batch_norm_context=[]
for bin_size in bin_size_list:
num_bins=int(1/bin_size)
for bin_index in range(num_bins):
min_bin=depth_gt_nan.nanquantile(bin_index*bin_size,dim=-1).unsqueeze(-1).unsqueeze(-1)
max_bin=depth_gt_nan.nanquantile((bin_index+1) * bin_size, dim=-1).unsqueeze(-1).unsqueeze(-1)
new_mask_valid=mask_valid
new_mask_valid=new_mask_valid & (depth_gt>=min_bin)
new_mask_valid = new_mask_valid & (depth_gt < max_bin)
batch_norm_context.append(new_mask_valid)
batch_norm_context = torch.stack(batch_norm_context, dim=0)
return batch_norm_context
def init_temp_masks_ds(level,image_size):
size=image_size
bin_size_list = [(1 / 2) ** (i) for i in range(level)]
bin_size_list.reverse()
map_level_list = []
for bin_size in bin_size_list: # e.g. 1/8
for h in range(int(1 / bin_size)):
for w in range(int(1 / bin_size)):
mask_new=torch.zeros(1,1,size,size)
mask_new[:,:, int(h * bin_size * size):int((h + 1) * bin_size * size),
int(w * bin_size * size):int((w + 1) * bin_size * size)] = 1
mask_new = mask_new> 0
map_level_list.append(mask_new)
batch_norm_context=torch.stack(map_level_list,dim=0)
return batch_norm_context
def get_contexts_ds( level, mask_valid):
templete_contexts=init_temp_masks_ds(level,mask_valid.shape[-1])
batch_norm_context = mask_valid.unsqueeze(0)
batch_norm_context = batch_norm_context.repeat(templete_contexts.shape[0], 1, 1, 1, 1)
batch_norm_context = batch_norm_context & templete_contexts
return batch_norm_context
class SSIMAE(nn.Module):
#modified from omnidata github https://github.com/EPFL-VILAB/omnidata
def __init__(self):
super().__init__()
def forward(self, depth_preds, depth_gt, mask_valid,dense):
depth_pred_aligned, depth_gt_aligned = masked_shift_and_scale(depth_preds, depth_gt, mask_valid) #normalize the depth maps
ssi_mae_loss = masked_l1_loss(depth_pred_aligned, depth_gt_aligned, mask_valid,dense)
return ssi_mae_loss
def compute_hdn_loss(SSI_LOSS,depth_preds,depth_gt,mask_valid_list):
hdn_loss_level = SSI_LOSS( #batch computation
depth_preds.unsqueeze(0).repeat(mask_valid_list.shape[0], 1, 1, 1, 1).reshape(-1,
*depth_preds.shape[
-3:]),
depth_gt.unsqueeze(0).repeat(mask_valid_list.shape[0], 1, 1, 1, 1).reshape(-1,
*depth_gt.shape[
-3:]),
mask_valid_list.reshape(-1, *mask_valid_list.shape[-3:]), dense=True)
hdn_loss_level_list = hdn_loss_level.reshape(*mask_valid_list.shape)
hdn_loss_level_list = hdn_loss_level_list.sum(dim=0) # summed loss generated by different contexts for all locations
mask_valid_list_times = mask_valid_list.sum(dim=0) # the number of contexts for each locations
valid_locations = (mask_valid_list_times != 0) # valid locations
hdn_loss_level_list[valid_locations] = hdn_loss_level_list[valid_locations] / mask_valid_list_times[
valid_locations] # mean loss in each location
hdn_loss = hdn_loss_level_list.sum() / mask_valid.sum() # average the losses of all locations
return hdn_loss
if __name__ == '__main__':
batch_size=2
depth_preds=torch.rand(2,1,384,384) #predicted depth maps
depth_gt=torch.rand(2,1,384,384) #ground truth depth maps
mask_valid=torch.rand(2,1,384,384)>0.5 #valid pixels
SSI_LOSS=SSIMAE() #ssi loss function
mask_valid_list_dr = get_contexts_dr(3, depth_gt, mask_valid) # get contexts by hdn_dr
mask_valid_list_dp = get_contexts_dp(3, depth_gt, mask_valid) # get contexts by hdn_dp
mask_valid_list_ds = get_contexts_ds(3, mask_valid) # get contexts by hdn_ds
loss_hdn_dr = compute_hdn_loss(SSI_LOSS, depth_preds, depth_gt, mask_valid_list_dr)
loss_hdn_dp = compute_hdn_loss(SSI_LOSS, depth_preds, depth_gt, mask_valid_list_dp)
loss_hdn_ds = compute_hdn_loss(SSI_LOSS, depth_preds, depth_gt, mask_valid_list_ds)
print (loss_hdn_ds,loss_hdn_dp,loss_hdn_dr)