-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathloss_functions.py
203 lines (154 loc) · 7.11 KB
/
loss_functions.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
193
194
195
196
197
198
199
200
201
202
203
from __future__ import division
import torch
from torch import nn, Tensor
import torch.nn.functional as F
from inverse_warp import inverse_warp, inverse_multiwarp
from inverse_warp import transform_module_pose_to_subaperture_pose
def multiwarp_photometric_loss(
tgt_lf, ref_lfs, intrinsics, depth, pose, metadata,
rotation_mode='euler', padding_mode='zeros'):
""" Computes photometric reconstruction loss across an entire lightfield.
Arguments:
tgt_lf: target lightfield (to reconstruct) -- [B, N, H, W]
ref_lfs: list of reference lightfields (to be sampled in reconstructing) - [[B, N, H, W]]
intrinsics: k matrix -- [B, 3, 3]
depth: [B, NCams, H, W]
pose: [B, SeqLen, 6]
metadata: metadata returned from the dataloader
rotation_mode: euler or quat
padding_mode: zeros or border
Returns:
photometric_warp_loss
warped_images
difference_images
"""
def one_scale(depth):
assert(pose.size(1) == len(ref_lfs))
reconstruction_loss = 0
b, n, h, w = depth.size()
downscale = tgt_lf.size(2)/h
tgt_lf_scaled = F.interpolate(tgt_lf, (h, w), mode='area')
ref_lf_scaled = [F.interpolate(ref_lf, (h, w), mode='area') for ref_lf in ref_lfs]
intrinsics_scaled = torch.cat((intrinsics[:, 0:2]/downscale, intrinsics[:, 2:]), dim=1)
warped_images = []
diff_images = []
# For every camera, perform the inverse warp
for N, cam in enumerate(metadata['cameras']):
current_depth = depth[:, N, :, :]
# For every reference image in the sequence
for i, ref_img in enumerate(ref_lf_scaled):
current_pose = transform_module_pose_to_subaperture_pose(pose[:, i], cam)
ref_img = ref_img[:, N:N+1, :, :]
ref_image_warped, valid_points = inverse_multiwarp(ref_img, current_depth, current_pose, intrinsics_scaled, rotation_mode, padding_mode)
diff = (tgt_lf_scaled[:, N:N+1, :, :] - ref_image_warped) * valid_points.unsqueeze(1).float()
reconstruction_loss += diff.abs().mean()
warped_images.append(ref_image_warped)
diff_images.append(diff)
return reconstruction_loss, warped_images, diff_images
if type(depth) not in [list, tuple]:
depth = [depth]
total_loss = 0
warped_image_results = []
difference_image_results = []
for d in depth:
loss, warped, diff = one_scale(d)
total_loss += loss
warped_image_results.append(warped)
difference_image_results.append(diff)
return total_loss, warped_image_results, difference_image_results
def photometric_reconstruction_loss(
tgt_img, ref_imgs, intrinsics,
depth, explainability_mask, pose,
rotation_mode='euler', padding_mode='zeros'):
def one_scale(depth, explainability_mask):
assert(explainability_mask is None or depth.size()[2:] == explainability_mask.size()[2:])
assert(pose.size(1) == len(ref_imgs))
reconstruction_loss = 0
b, _, h, w = depth.size()
downscale = tgt_img.size(2)/h
tgt_img_scaled = F.interpolate(tgt_img, (h, w), mode='area')
ref_imgs_scaled = [F.interpolate(ref_img, (h, w), mode='area') for ref_img in ref_imgs]
intrinsics_scaled = torch.cat((intrinsics[:, 0:2]/downscale, intrinsics[:, 2:]), dim=1)
warped_imgs = []
diff_maps = []
for i, ref_img in enumerate(ref_imgs_scaled):
current_pose = pose[:, i]
ref_img_warped, valid_points = inverse_warp(ref_img, depth[:,0], current_pose, intrinsics_scaled, rotation_mode, padding_mode)
diff = (tgt_img_scaled - ref_img_warped) * valid_points.unsqueeze(1).float()
if explainability_mask is not None:
diff = diff * explainability_mask[:,i:i+1].expand_as(diff)
reconstruction_loss += diff.abs().mean()
assert((reconstruction_loss == reconstruction_loss).item() == 1)
warped_imgs.append(ref_img_warped[0])
diff_maps.append(diff[0])
return reconstruction_loss, warped_imgs, diff_maps
warped_results, diff_results = [], []
if type(explainability_mask) not in [tuple, list]:
explainability_mask = [explainability_mask]
if type(depth) not in [list, tuple]:
depth = [depth]
total_loss = 0
for d, mask in zip(depth, explainability_mask):
loss, warped, diff = one_scale(d, mask)
total_loss += loss
warped_results.append(warped)
diff_results.append(diff)
return total_loss, warped_results, diff_results
def explainability_loss(mask):
if type(mask) not in [tuple, list]:
mask = [mask]
loss = 0
for mask_scaled in mask:
ones_var = torch.ones_like(mask_scaled)
loss += nn.functional.binary_cross_entropy(mask_scaled, ones_var)
return loss
def smooth_loss(pred_map):
def gradient(pred):
D_dy = pred[:, :, 1:] - pred[:, :, :-1]
D_dx = pred[:, :, :, 1:] - pred[:, :, :, :-1]
return D_dx, D_dy
if type(pred_map) not in [tuple, list]:
pred_map = [pred_map]
loss = 0
weight = 1.
for scaled_map in pred_map:
dx, dy = gradient(scaled_map)
dx2, dxdy = gradient(dx)
dydx, dy2 = gradient(dy)
loss += (dx2.abs().mean() + dxdy.abs().mean() + dydx.abs().mean() + dy2.abs().mean())*weight
weight /= 2.3 # don't ask me why it works better
return loss
def pose_loss(pose, pose_gt):
pred_pose_magnitude = pose[:, :, :3].norm(dim=2)
pose_gt_magnitude = pose_gt[:, :, :3].norm(dim=2)
error = (pred_pose_magnitude - pose_gt_magnitude).abs().mean()
return error
@torch.no_grad()
def compute_errors(gt, pred, crop=True):
abs_diff, abs_rel, sq_rel, a1, a2, a3 = 0,0,0,0,0,0
batch_size = gt.size(0)
'''
crop used by Garg ECCV16 to reprocude Eigen NIPS14 results
construct a mask of False values, with the same size as target
and then set to True values inside the crop
'''
if crop:
crop_mask = gt[0] != gt[0]
y1,y2 = int(0.40810811 * gt.size(1)), int(0.99189189 * gt.size(1))
x1,x2 = int(0.03594771 * gt.size(2)), int(0.96405229 * gt.size(2))
crop_mask[y1:y2,x1:x2] = 1
for current_gt, current_pred in zip(gt, pred):
valid = (current_gt > 0) & (current_gt < 80)
if crop:
valid = valid & crop_mask
valid_gt = current_gt[valid]
valid_pred = current_pred[valid].clamp(1e-3, 80)
valid_pred = valid_pred * torch.median(valid_gt)/torch.median(valid_pred)
thresh = torch.max((valid_gt / valid_pred), (valid_pred / valid_gt))
a1 += (thresh < 1.25).float().mean()
a2 += (thresh < 1.25 ** 2).float().mean()
a3 += (thresh < 1.25 ** 3).float().mean()
abs_diff += torch.mean(torch.abs(valid_gt - valid_pred))
abs_rel += torch.mean(torch.abs(valid_gt - valid_pred) / valid_gt)
sq_rel += torch.mean(((valid_gt - valid_pred)**2) / valid_gt)
return [metric.item() / batch_size for metric in [abs_diff, abs_rel, sq_rel, a1, a2, a3]]