Spaces:
Sleeping
Sleeping
File size: 7,907 Bytes
ec9a6bc |
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 torch
import numpy as np
from einops import rearrange
class Calibrator():
def __init__(self, dataset, bfm, camera, recorder):
self.dataset = dataset
self.bfm = bfm
self.camera = camera
self.recorder = recorder
self.device = torch.device('cuda:0')
self.optimizer1 = torch.optim.Adam([{'params' : self.bfm.parameters(), 'lr' : 1e-2}])
self.optimizer2 = torch.optim.Adam([{'params' : self.bfm.parameters(), 'lr' : 1e-3}])
def calibrate(self):
landmarks_gt, extrinsics0, intrinsics0, eids = self.dataset.get_item()
landmarks_gt = torch.from_numpy(landmarks_gt).float().to(self.device)
extrinsics0 = torch.from_numpy(extrinsics0).float().to(self.device)
intrinsics0 = torch.from_numpy(intrinsics0).float().to(self.device)
extrinsics = rearrange(extrinsics0, 'b v x y -> (b v) x y')
intrinsics = rearrange(intrinsics0, 'b v x y -> (b v) x y')
pprev_loss = 1e8
prev_loss = 1e8
for i in range(100000000):
self.optimizer1.zero_grad()
_, landmarks_3d = self.bfm()
landmarks_3d = landmarks_3d.unsqueeze(1).repeat(1, landmarks_gt.shape[1], 1, 1)
landmarks_3d = rearrange(landmarks_3d, 'b v x y -> (b v) x y')
landmarks_2d = self.project(landmarks_3d, intrinsics, extrinsics)
landmarks_2d = rearrange(landmarks_2d, '(b v) x y -> b v x y', b=landmarks_gt.shape[0])
pro_loss = (((landmarks_2d / self.camera.image_size - landmarks_gt[:, :, :, 0:2] / self.camera.image_size) * landmarks_gt[:, :, :, 2:3]) ** 2).sum(-1).sum(-2).mean()
reg_loss = self.bfm.reg_loss(5e-6, 1e-6)
loss = pro_loss + reg_loss
loss.backward()
self.optimizer1.step()
if abs(loss.item() - prev_loss) < 1e-8 and abs(loss.item() - pprev_loss) < 1e-7:
if i % 100 == 0:
print(pro_loss.item(), reg_loss.item())
break
else:
pprev_loss = prev_loss
prev_loss = loss.item()
if i % 100 == 0:
print(pro_loss.item(), reg_loss.item())
for i in range(100000000):
self.optimizer2.zero_grad()
_, landmarks_3d = self.bfm()
landmarks_3d = landmarks_3d.unsqueeze(1).repeat(1, landmarks_gt.shape[1], 1, 1)
landmarks_3d = rearrange(landmarks_3d, 'b v x y -> (b v) x y')
landmarks_2d = self.project(landmarks_3d, intrinsics, extrinsics)
landmarks_2d = rearrange(landmarks_2d, '(b v) x y -> b v x y', b=landmarks_gt.shape[0])
pro_loss = (((landmarks_2d / self.camera.image_size - landmarks_gt[:, :, :, 0:2] / self.camera.image_size) * landmarks_gt[:, :, :, 2:3]) ** 2).sum(-1).sum(-2).mean()
reg_loss = self.bfm.reg_loss(5e-6, 1e-6)
loss = pro_loss + reg_loss
loss.backward()
self.optimizer2.step()
if abs(loss.item() - prev_loss) < 1e-11 and abs(loss.item() - pprev_loss) < 1e-10:
if i % 100 == 0:
print(pro_loss.item(), reg_loss.item())
break
else:
pprev_loss = prev_loss
prev_loss = loss.item()
if i % 100 == 0:
print(pro_loss.item(), reg_loss.item())
log = {
'eids': eids,
'landmarks_gt': landmarks_gt,
'landmarks_2d': landmarks_2d.detach(),
'bfm': self.bfm,
'intrinsics': intrinsics0,
'extrinsics': extrinsics0
}
self.recorder.log(log)
def project(self, points_3d, intrinsic, extrinsic):
points_3d = points_3d.permute(0,2,1)
calibrations = torch.bmm(intrinsic, extrinsic)
points_2d = self.camera.perspective(points_3d, calibrations)
points_2d = points_2d.permute(0,2,1)
return points_2d
class CalibratorSingleView():
def __init__(self, dataset, bfm, camera, recorder):
self.dataset = dataset
self.bfm = bfm
self.camera = camera
self.recorder = recorder
self.device = torch.device('cuda:0')
self.optimizer1 = torch.optim.Adam([{'params' : self.bfm.parameters(), 'lr' : 1e-2}])
self.optimizer2 = torch.optim.Adam([{'params' : self.bfm.parameters(), 'lr' : 1e-3}])
def calibrate(self):
landmarks_gt, extrinsics0, intrinsics0, eids = self.dataset.get_item()
landmarks_gt = torch.from_numpy(landmarks_gt).float().to(self.device)
extrinsics0 = torch.from_numpy(extrinsics0).float().to(self.device)
intrinsics0 = torch.from_numpy(intrinsics0).float().to(self.device)
extrinsics = rearrange(extrinsics0, 'b v x y -> (b v) x y')
intrinsics = rearrange(intrinsics0, 'b v x y -> (b v) x y')
pprev_loss = 1e8
prev_loss = 1e8
for i in range(3000):
self.optimizer1.zero_grad()
_, landmarks_3d = self.bfm()
landmarks_3d = landmarks_3d.unsqueeze(1).repeat(1, landmarks_gt.shape[1], 1, 1)
landmarks_3d = rearrange(landmarks_3d, 'b v x y -> (b v) x y')
landmarks_2d = self.project(landmarks_3d, intrinsics, extrinsics)
landmarks_2d = rearrange(landmarks_2d, '(b v) x y -> b v x y', b=landmarks_gt.shape[0])
pro_loss = (((landmarks_2d / self.camera.image_size - landmarks_gt[:, :, :, 0:2] / self.camera.image_size) * landmarks_gt[:, :, :, 2:3]) ** 2).sum(-1).sum(-2).mean()
reg_loss = self.bfm.reg_loss(5e-6, 1e-6)
loss = pro_loss + reg_loss
loss.backward()
self.optimizer1.step()
if False: #abs(loss.item() - prev_loss) < 1e-8 and abs(loss.item() - pprev_loss) < 1e-7:
print(prev_loss, 'optimization ends')
break
else:
pprev_loss = prev_loss
prev_loss = loss.item()
if i % 100 == 0:
print(i, prev_loss)
for i in range(1000):
self.optimizer2.zero_grad()
_, landmarks_3d = self.bfm()
landmarks_3d = landmarks_3d.unsqueeze(1).repeat(1, landmarks_gt.shape[1], 1, 1)
landmarks_3d = rearrange(landmarks_3d, 'b v x y -> (b v) x y')
landmarks_2d = self.project(landmarks_3d, intrinsics, extrinsics)
landmarks_2d = rearrange(landmarks_2d, '(b v) x y -> b v x y', b=landmarks_gt.shape[0])
pro_loss = (((landmarks_2d / self.camera.image_size - landmarks_gt[:, :, :, 0:2] / self.camera.image_size) * landmarks_gt[:, :, :, 2:3]) ** 2).sum(-1).sum(-2).mean()
reg_loss = self.bfm.reg_loss(5e-6, 1e-6) + self.bfm.temporal_smooth_loss(3e-5)
loss = pro_loss + reg_loss
loss.backward()
self.optimizer2.step()
if False: #abs(loss.item() - prev_loss) < 1e-11 and abs(loss.item() - pprev_loss) < 1e-10:
print(prev_loss, 'optimization ends')
break
else:
pprev_loss = prev_loss
prev_loss = loss.item()
if i % 100 == 0:
print(i, prev_loss)
log = {
'eids': eids,
'landmarks_gt': landmarks_gt,
'landmarks_2d': landmarks_2d.detach(),
'bfm': self.bfm,
'intrinsics': intrinsics0,
'extrinsics': extrinsics0
}
self.recorder.log(log)
def project(self, points_3d, intrinsic, extrinsic):
points_3d = points_3d.permute(0,2,1)
calibrations = torch.bmm(intrinsic, extrinsic)
points_2d = self.camera.perspective(points_3d, calibrations)
points_2d = points_2d.permute(0,2,1)
return points_2d
|