File size: 7,753 Bytes
f06f310 |
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 |
import re
import os
import sys
import logging
import numpy as np
from collections import OrderedDict
# logging.basicConfig(level=logging.INFO,
# format='%(asctime)s %(levelname)-8s [%(filename)s:%(lineno)d] %(message)s',)
import torch
import torch.nn as nn
import torch.nn.functional as F
from core.utils.plane import convert2patch
class Geometry_MLP(nn.Module):
def __init__(self, args):
super(Geometry_MLP, self).__init__()
self.args = args
self.reg = nn.Sequential(
nn.Linear(3,3),
nn.Linear(3,2),
)
if args.geo_fusion.lower()=="max":
self.fusion = nn.AdaptiveMaxPool1d(1)
elif args.geo_fusion.lower()=="mean":
self.fusion = nn.AdaptiveAvgPool1d(1)
else:
raise Exception(f"{args.geo_fusion} is not supported")
def forward(self, img_coord, flow_up):
# (1,4,factor*factor,H,W)
factor = 2 ** self.args.n_downsample
fit_points = torch.cat([img_coord, flow_up], dim=1)
fit_points = convert2patch(fit_points,
patch_size=factor,
div_last=False) # (1,3,factor*factor,H,W)
A = fit_points[:,:3].permute((0,2,3,4,1)) # (1,factor*factor,H,W,3)
ab_proposals = self.reg(A) # (1,factor*factor,H,W,2)
B,L,H,W,C = ab_proposals.shape
ab = self.fusion(ab_proposals.view(B,L,-1).transpose(-1,-2)) # (1,H*W*2,1)
ab = ab.view(B,H,W,C).permute((0,3,1,2)) # (1,2,H,W)
geo = torch.cat([disparity[:,:1],ab], dim=1)
return ab
class Geometry_Conv(nn.Module):
def __init__(self, args):
super(Geometry_Conv, self).__init__()
self.args = args
self.reg = nn.Sequential(
nn.Conv2d(3, 4, kernel_size=3, padding=1, stride=1),
nn.LeakyReLU(inplace=True),
nn.Conv2d(4, 8, kernel_size=3, padding=1, stride=2),
nn.LeakyReLU(inplace=True),
nn.Conv2d(8, 5, kernel_size=3, padding=1, stride=2),
nn.LeakyReLU(inplace=True),
nn.Conv2d(5, 5, kernel_size=1, padding=0, stride=1),
)
def forward(self, img_coord, disparity_up, disparity):
# img_coord: (1,2,H*factor,W*factor)
# disparity_up: (1,1,H*factor,W*factor)
# disparity: (1,1,H,W)
# factor = 2 ** self.args.n_downsample
# points = torch.cat([img_coord, disparity_up], dim=1) # (1,3,factor*H,factor*W)
points = torch.cat([img_coord, disparity_up.detach()], dim=1) # (1,3,factor*H,factor*W)
rest_params = self.reg(points) # (1,5,H,W)
params = torch.cat([disparity,rest_params], dim=1) # (1,6,H,W)
return params
class Geometry_Conv_Split(nn.Module):
def __init__(self, args):
super(Geometry_Conv_Split, self).__init__()
self.args = args
self.encode = nn.Sequential(
nn.Conv2d(3, 4, kernel_size=3, padding=1, stride=1),
nn.LeakyReLU(inplace=True),
nn.Conv2d(4, 8, kernel_size=3, padding=1, stride=2),
nn.LeakyReLU(inplace=True),
)
self.decode_plane = nn.Sequential(
nn.Conv2d(8, 4, kernel_size=3, padding=1, stride=2),
nn.LeakyReLU(inplace=True),
nn.Conv2d(4, 2, kernel_size=1, padding=0, stride=1),
)
self.decode_curvature = nn.Sequential(
nn.Conv2d(8, 4, kernel_size=3, padding=1, stride=2),
nn.LeakyReLU(inplace=True),
nn.Conv2d(4, 3, kernel_size=1, padding=0, stride=1),
)
def forward(self, img_coord, disparity_up, disparity):
# img_coord: (1,2,H*factor,W*factor)
# disparity_up: (1,1,H*factor,W*factor)
# disparity: (1,1,H,W)
# factor = 2 ** self.args.n_downsample
points = torch.cat([img_coord, disparity_up], dim=1) # (1,3,factor*H,factor*W)
latten = self.encode(points) # (1,8,factor*H/2,factor*W/2)
plane_ab = self.decode_plane(latten) # (1,2,H,W)
hessian_g = self.decode_curvature(latten) # (1,3,H,W)
params = torch.cat([disparity,plane_ab,hessian_g], dim=1) # (1,6,H,W)
return params
class LBPEncoder(nn.Module):
"""
Computes the modified Local Binary Patterns (LBP) of an image using custom neighbor offsets.
"""
def __init__(self, args):
super(LBPEncoder, self).__init__()
self.args = args
self.lbp_neighbor_offsets = self._parse_offsets(self.args.lbp_neighbor_offsets)
self._build_lbp_kernel()
self.sigmoid = nn.Sigmoid()
def _build_lbp_kernel(self):
# Determine the kernel size based on the maximum offset
self.num_neighbors = len(self.lbp_neighbor_offsets)
self.max_offset = int(np.abs(self.lbp_neighbor_offsets).max())
self.kernel_size = 2 * self.max_offset + 1
self.padding = self.max_offset
# Initialize the convolution layer for depthwise convolution
self.lbp_conv = nn.Conv2d(
in_channels=1,
out_channels=self.num_neighbors,
kernel_size=self.kernel_size,
padding=self.padding,
padding_mode="replicate",
bias=False,
groups=1 # Since in_channels=1, groups=1 makes it depthwise
)
self.lbp_weight = torch.zeros(self.num_neighbors, 1,
self.kernel_size, self.kernel_size).float()
center_y, center_x = self.max_offset, self.max_offset
for idx, (dy, dx) in enumerate(self.lbp_neighbor_offsets):
# Compute the position in the kernel for the neighbor
y, x = center_y + dy, center_x + dx
if 0 <= y < self.kernel_size and 0 <= x < self.kernel_size:
self.lbp_weight[idx, 0, y, x] = 1.0
self.lbp_weight[idx, 0, center_y, center_x] = -1.0
else:
raise ValueError(f"Offset ({dy}, {dx}) is out of kernel bounds.")
# Assign the weight to the convolution layer
self.lbp_conv.weight = nn.Parameter(self.lbp_weight)
self.lbp_conv.weight.requires_grad = False # Ensure weights are not updated during training
def _parse_offsets(self, offsets_str):
"""
Parses a string to extract neighbor offsets.
Parameters:
offsets_str (str): String defining neighbor offsets, e.g., "(-1,-1), (1,1), (-1,1), (1,-1)"
Returns:
list of tuples: List of neighbor offsets.
"""
# extract coordinate pairs
pattern = r'\((-?\d+),\s*(-?\d+)\)'
matches = re.findall(pattern, offsets_str)
if not matches:
raise ValueError(offsets_str + ": not suppoted format, please check it!")
offsets = [(int(y), int(x)) for y, x in matches]
return np.array(offsets)
def forward(self, img):
"""
Parameters:
img (torch.Tensor): Grayscale image tensor of shape [N, 1, H, W].
Returns:
torch.Tensor: Modified LBP image of shape [N, C, H, W].
"""
with torch.no_grad():
# Apply convolution to compute differences directly
differences = self.lbp_conv(img) # Shape: [1, N, H, W] due to padding
# Apply sigmoid to the differences to get encoding values between 0 and 1
encoding = self.sigmoid(differences) # Shape: [1, N, H, W]
return encoding
|