File size: 19,442 Bytes
91126af |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 |
# Copyright (C) 2024-present Naver Corporation. All rights reserved.
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
#
# --------------------------------------------------------
# MASt3R heads
# --------------------------------------------------------
import torch
import torch.nn.functional as F
import mast3r.utils.path_to_dust3r # noqa
from dust3r.heads.postprocess import reg_dense_depth, reg_dense_conf # noqa
from dust3r.heads.dpt_head import PixelwiseTaskWithDPT # noqa
import dust3r.utils.path_to_croco # noqa
from models.blocks import Mlp # noqa
import torch.nn as nn
def reg_desc(desc, mode):
if 'norm' in mode:
desc = desc / desc.norm(dim=-1, keepdim=True)
else:
raise ValueError(f"Unknown desc mode {mode}")
return desc
def postprocess(out, depth_mode, conf_mode, desc_dim=None, desc_mode='norm', two_confs=False, desc_conf_mode=None):
if desc_conf_mode is None:
desc_conf_mode = conf_mode
fmap = out.permute(0, 2, 3, 1) # B,H,W,D
res = dict(pts3d=reg_dense_depth(fmap[..., 0:3], mode=depth_mode))
if conf_mode is not None:
res['conf'] = reg_dense_conf(fmap[..., 3], mode=conf_mode)
if desc_dim is not None:
start = 3 + int(conf_mode is not None)
res['desc'] = fmap[..., start:]
# if two_confs:
# res['desc_conf'] = reg_dense_conf(fmap[..., start + desc_dim], mode=desc_conf_mode)
# else:
# res['desc_conf'] = res['conf'].clone()
return res
class Cat_MLP_LocalFeatures_DPT_Pts3d(PixelwiseTaskWithDPT):
""" Mixture between MLP and DPT head that outputs 3d points and local features (with MLP).
The input for both heads is a concatenation of Encoder and Decoder outputs
"""
def __init__(self, net, has_conf=False, local_feat_dim=16, hidden_dim_factor=4., hooks_idx=None, dim_tokens=None,
num_channels=1, postprocess=None, feature_dim=256, last_dim=32, depth_mode=None, conf_mode=None, head_type="regression", **kwargs):
super().__init__(num_channels=num_channels, feature_dim=feature_dim, last_dim=last_dim, hooks_idx=hooks_idx,
dim_tokens=dim_tokens, depth_mode=depth_mode, postprocess=postprocess, conf_mode=conf_mode, head_type=head_type)
self.local_feat_dim = local_feat_dim
patch_size = net.patch_embed.patch_size
if isinstance(patch_size, tuple):
assert len(patch_size) == 2 and isinstance(patch_size[0], int) and isinstance(
patch_size[1], int), "What is your patchsize format? Expected a single int or a tuple of two ints."
assert patch_size[0] == patch_size[1], "Error, non square patches not managed"
patch_size = patch_size[0]
self.patch_size = patch_size
self.desc_mode = net.desc_mode
self.has_conf = has_conf
self.two_confs = net.two_confs # independent confs for 3D regr and descs
self.desc_conf_mode = net.desc_conf_mode
idim = net.enc_embed_dim + net.dec_embed_dim
self.head_local_features = Mlp(in_features=idim,
hidden_features=int(hidden_dim_factor * idim),
out_features=(self.local_feat_dim + self.two_confs) * self.patch_size**2)
def forward(self, decout, img_shape):
# pass through the heads
pts3d = self.dpt(decout, image_size=(img_shape[0], img_shape[1]))
# recover encoder and decoder outputs
enc_output, dec_output = decout[0], decout[-1]
cat_output = torch.cat([enc_output, dec_output], dim=-1) # concatenate
H, W = img_shape
B, S, D = cat_output.shape
# extract local_features
local_features = self.head_local_features(cat_output) # B,S,D
local_features = local_features.transpose(-1, -2).view(B, -1, H // self.patch_size, W // self.patch_size)
local_features = F.pixel_shuffle(local_features, self.patch_size) # B,d,H,W
# post process 3D pts, descriptors and confidences
out = torch.cat([pts3d, local_features], dim=1)
if self.postprocess:
out = self.postprocess(out,
depth_mode=self.depth_mode,
conf_mode=self.conf_mode,
desc_dim=self.local_feat_dim,
desc_mode=self.desc_mode,
two_confs=self.two_confs,
desc_conf_mode=self.desc_conf_mode)
# out.update({'local_token': local_token})
return out
class DPT_depth(PixelwiseTaskWithDPT):
""" Mixture between MLP and DPT head that outputs 3d points and local features (with MLP).
The input for both heads is a concatenation of Encoder and Decoder outputs
"""
def __init__(self, net, has_conf=False, hidden_dim_factor=4., hooks_idx=None, dim_tokens=None,
num_channels=1, postprocess=None, feature_dim=256, last_dim=32, depth_mode=None, conf_mode=None, head_type="regression", **kwargs):
super().__init__(num_channels=num_channels, feature_dim=feature_dim, last_dim=last_dim, hooks_idx=hooks_idx,
dim_tokens=dim_tokens, depth_mode=depth_mode, postprocess=postprocess, conf_mode=conf_mode, head_type=head_type)
patch_size = net.patch_embed.patch_size
if isinstance(patch_size, tuple):
assert len(patch_size) == 2 and isinstance(patch_size[0], int) and isinstance(
patch_size[1], int), "What is your patchsize format? Expected a single int or a tuple of two ints."
assert patch_size[0] == patch_size[1], "Error, non square patches not managed"
patch_size = patch_size[0]
self.patch_size = patch_size
self.desc_mode = net.desc_mode
self.has_conf = has_conf
self.two_confs = net.two_confs # independent confs for 3D regr and descs
self.desc_conf_mode = net.desc_conf_mode
idim = net.enc_embed_dim + net.dec_embed_dim
# self.conf_mode = conf_mode
def forward(self, decout, img_shape):
# pass through the heads
pts3d = self.dpt(decout, image_size=(img_shape[0], img_shape[1]))
out = pts3d
# post process 3D pts, descriptors and confidences
# out = torch.cat([pts3d, local_features], dim=1)
fmap = out.permute(0, 2, 3, 1) # B,H,W,3
res = {}
res['depth'] = torch.exp(fmap[...,:1]-1).clamp(0.0001, 1000.)
# res['depth_scaling'] = fmap[...,1:4]
res['depth_conf'] = reg_dense_conf(fmap[..., -1:], mode=self.conf_mode)
res['desc'] = fmap[..., 1:]
# out.update({'local_token': local_token})
return res
class Cat_GS_LocalFeatures_DPT_Pts3d(PixelwiseTaskWithDPT):
""" Mixture between MLP and DPT head that outputs 3d points and local features (with MLP).
The input for both heads is a concatenation of Encoder and Decoder outputs
"""
def __init__(self, net, has_conf=False, local_feat_dim=16, hidden_dim_factor=4., hooks_idx=None, dim_tokens=None,
num_channels=1, postprocess=None, feature_dim=256, last_dim=32, depth_mode=None, conf_mode=None, head_type="regression", **kwargs):
super().__init__(num_channels=num_channels, feature_dim=feature_dim, last_dim=last_dim, hooks_idx=hooks_idx,
dim_tokens=dim_tokens, depth_mode=depth_mode, postprocess=postprocess, conf_mode=conf_mode, head_type=head_type)
self.local_feat_dim = local_feat_dim
patch_size = net.patch_embed.patch_size
if isinstance(patch_size, tuple):
assert len(patch_size) == 2 and isinstance(patch_size[0], int) and isinstance(
patch_size[1], int), "What is your patchsize format? Expected a single int or a tuple of two ints."
assert patch_size[0] == patch_size[1], "Error, non square patches not managed"
patch_size = patch_size[0]
self.patch_size = patch_size
self.desc_mode = net.desc_mode
self.has_conf = has_conf
self.two_confs = net.two_confs # independent confs for 3D regr and descs
self.desc_conf_mode = net.desc_conf_mode
idim = net.enc_embed_dim + net.dec_embed_dim
def forward(self, decout, img_shape):
# pass through the heads
out = self.dpt(decout, image_size=(img_shape[0], img_shape[1]))
# recover encoder and decoder outputs
# enc_output, dec_output = decout[0], decout[-1]
# cat_output = torch.cat([enc_output, dec_output], dim=-1) # concatenate
# H, W = img_shape
# B, S, D = cat_output.shape
# post process 3D pts, descriptors and confidences
# out = torch.cat([pts3d, local_features], dim=1)
if self.postprocess:
out = self.postprocess(out,
depth_mode=self.depth_mode,
conf_mode=self.conf_mode,
desc_dim=self.local_feat_dim,
desc_mode=self.desc_mode,
two_confs=self.two_confs,
desc_conf_mode=self.desc_conf_mode)
# out.update({'local_token': local_token})
return out
class UNet(nn.Module):
def __init__(self, in_channels, out_channels, hidden_dim):
super(UNet, self).__init__()
# ηΌη ε¨
self.enc1 = self.conv_block(in_channels, hidden_dim)
# self.downsample = nn.Conv2d(hidden_dim, hidden_dim * 2, kernel_size=2, stride=2) # δΈιζ ·
# 解η ε¨
# self.dec1 = self.upconv_block(hidden_dim * 2, hidden_dim)
self.dec2 = nn.Conv2d(hidden_dim, out_channels, kernel_size=3, padding=1)
def conv_block(self, in_channels, out_channels):
return nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
nn.GELU(),
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
nn.GELU()
)
def upconv_block(self, in_channels, out_channels):
return nn.Sequential(
nn.ConvTranspose2d(in_channels, out_channels, kernel_size=2, stride=2),
nn.GELU()
)
def forward(self, x):
# ηΌη
enc1 = self.enc1(x)
dec2 = self.dec2(enc1)
return dec2
class gs_head_heavy(nn.Module):
def __init__(self,
feature_dim,
last_dim,
high_feature,
sh_degree = 2,
):
super().__init__()
self.high_feature = high_feature
self.high_feature_fusion = UNet(high_feature, high_feature, high_feature)
sh_degree = sh_degree
self.feat_sh = nn.Sequential(
nn.Conv2d(feature_dim, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
)
self.color = nn.Conv2d(last_dim, 3, kernel_size=1, stride=1, padding=0)
self.sh_high_fre = nn.Conv2d(last_dim, (sh_degree + 1) ** 2 * 3 - 3, kernel_size=1, stride=1, padding=0)
self.feat_opacity = nn.Sequential(
nn.Conv2d(feature_dim, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
nn.Conv2d(last_dim, 1, kernel_size=1, stride=1, padding=0)
)
self.feat_scaling = nn.Sequential(
nn.Conv2d(high_feature, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
nn.Conv2d(last_dim, 3, kernel_size=1, stride=1, padding=0)
)
self.feat_rotation = nn.Sequential(
nn.Conv2d(high_feature, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
nn.Conv2d(last_dim, 4, kernel_size=1, stride=1, padding=0)
)
self.feat_scaling[-1].weight.data.normal_(mean=0, std=0.02)
self.feat_opacity[-1].weight.data.normal_(mean=0, std=0.02)
self.color.weight.data.normal_(mean=0, std=0.88)
self.sh_high_fre.weight.data.normal_(mean=0, std=0.02)
def forward(self, x, true_shape):
# H, W = x.shape[-2:]
# if H != H_org or W != W_org:
# x = x.permute(0, 1, 3, 2)
x = x[0]
x = x.permute(0,3,1,2) # B,H,W,D
assert x.shape[-1] == true_shape[-1]
high_feature = self.high_feature_fusion(x[:, :self.high_feature])
fusion_feature = torch.cat([high_feature, x[:, self.high_feature:]], dim=1)
feat_opacity = self.feat_opacity(fusion_feature)
feat_scaling = self.feat_scaling(high_feature)
feat_rotation = self.feat_rotation(high_feature)
featuresh = self.feat_sh(fusion_feature)
feat_color = self.color(featuresh)
feat_sh = self.sh_high_fre(featuresh)
feat_feature = torch.cat([feat_color, feat_sh], dim=1)
feat_feature = feat_feature.permute(0, 2, 3, 1) # B,H,W,3
feat_opacity = feat_opacity - 2
feat_opacity = feat_opacity.permute(0, 2, 3, 1) # B,H,W,1
feat_scaling = feat_scaling.permute(0, 2, 3, 1) # B,H,W,1
feat_rotation = feat_rotation.permute(0, 2, 3, 1) # B,H,W,1
res = dict(feature=feat_feature, opacity=feat_opacity, scaling=feat_scaling, rotation=feat_rotation)
return res
class gs_head(nn.Module):
def __init__(self,
feature_dim,
last_dim,
high_feature,
):
super().__init__()
self.high_feature = high_feature
self.feat_feature = nn.Sequential(
nn.Conv2d(feature_dim, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
nn.Conv2d(last_dim, 3, kernel_size=1, stride=1, padding=0)
)
self.feat_opacity = nn.Sequential(
nn.Conv2d(feature_dim, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
nn.Conv2d(last_dim, 1, kernel_size=1, stride=1, padding=0)
)
self.feat_scaling = nn.Sequential(
nn.Conv2d(high_feature, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
nn.Conv2d(last_dim, 3, kernel_size=1, stride=1, padding=0)
)
self.feat_rotation = nn.Sequential(
nn.Conv2d(high_feature, last_dim, kernel_size=3, stride=1, padding=1),
nn.GELU(),
nn.Conv2d(last_dim, 4, kernel_size=1, stride=1, padding=0)
)
self.feat_scaling[-1].weight.data.normal_(mean=0, std=0.02)
self.feat_opacity[-1].weight.data.normal_(mean=0, std=0.02)
self.feat_feature[-1].weight.data.normal_(mean=0, std=0.5)
def forward(self, x, true_shape):
# H, W = x.shape[-2:]
# if H != H_org or W != W_org:
# x = x.permute(0, 1, 3, 2)
x = x[0]
x = x.permute(0,3,1,2) # B,H,W,D
assert x.shape[-1] == true_shape[-1]
feat_opacity = self.feat_opacity(x)
feat_scaling = self.feat_scaling(x[:, :self.high_feature])
feat_rotation = self.feat_rotation(x[:, :self.high_feature])
feat_feature = self.feat_feature(x)
feat_feature = feat_feature.permute(0, 2, 3, 1) # B,H,W,3
feat_opacity = feat_opacity - 2
feat_opacity = feat_opacity.permute(0, 2, 3, 1) # B,H,W,1
feat_scaling = feat_scaling.permute(0, 2, 3, 1) # B,H,W,1
feat_rotation = feat_rotation.permute(0, 2, 3, 1) # B,H,W,1
res = dict(feature=feat_feature, opacity=feat_opacity, scaling=feat_scaling, rotation=feat_rotation)
return res
def mast3r_head_factory(head_type, output_mode, net, has_conf=False, sh_degree=2):
"""" build a prediction head for the decoder
"""
if head_type == 'catmlp+dpt' and output_mode.startswith('pts3d+desc'):
local_feat_dim = int(output_mode[10:])
assert net.dec_depth > 9
l2 = net.dec_depth
feature_dim = 256
last_dim = feature_dim // 2
out_nchan = 3
ed = net.enc_embed_dim
dd = net.dec_embed_dim
return Cat_MLP_LocalFeatures_DPT_Pts3d(net, local_feat_dim=local_feat_dim, has_conf=has_conf,
num_channels=out_nchan + has_conf,
feature_dim=feature_dim,
last_dim=last_dim,
hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],
dim_tokens=[ed, dd, dd, dd],
postprocess=postprocess,
depth_mode=net.depth_mode,
conf_mode=net.conf_mode,
head_type='regression')
elif output_mode=='depth_conf_scaling':
local_feat_dim = 24
assert net.dec_depth > 9
l2 = net.dec_depth
feature_dim = 256
last_dim = feature_dim // 2
out_nchan = 1
ed = net.enc_embed_dim
dd = net.dec_embed_dim
return DPT_depth(net, has_conf=has_conf,
num_channels=out_nchan + local_feat_dim + net.two_confs,
feature_dim=feature_dim,
last_dim=last_dim,
hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],
dim_tokens=[ed, dd, dd, dd],
postprocess=postprocess,
depth_mode=net.depth_mode,
conf_mode=net.conf_mode,
head_type='regression')
elif head_type == 'dpt_gs' and output_mode.startswith('pts3d+desc'):
local_feat_dim = int(output_mode[10:])
assert net.dec_depth > 9
l2 = net.dec_depth
feature_dim = 256
last_dim = feature_dim // 2
out_nchan = 3
ed = net.enc_embed_dim
dd = net.dec_embed_dim
return Cat_GS_LocalFeatures_DPT_Pts3d(net, has_conf=has_conf,
num_channels=out_nchan + has_conf + local_feat_dim + net.two_confs,
feature_dim=feature_dim,
last_dim=last_dim,
hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],
dim_tokens=[ed, dd, dd, dd],
postprocess=postprocess,
depth_mode=net.depth_mode,
conf_mode=net.conf_mode,
head_type='regression')
elif head_type == 'gs':
local_feat_dim = int(output_mode[10:]) + 1 + 16
return gs_head(feature_dim=local_feat_dim, last_dim=local_feat_dim//2, high_feature=int(output_mode[10:]) + 1)
elif head_type == 'sh':
local_feat_dim = int(output_mode[10:]) + 1 + 16
return gs_head_heavy(feature_dim=local_feat_dim, last_dim=local_feat_dim//2, high_feature=int(output_mode[10:]) + 1, sh_degree=sh_degree)
else:
raise NotImplementedError(
f"unexpected {head_type=} and {output_mode=}")
|