1.前置
(1)像素深度
像素深度指存储每个像素用的位数,像素深度越深,所占用存储空间越大;像素深度越浅,图像越粗糙和不自然。
(假设一幅彩色图像的每个像素用R,G,B三个分量表示,若每个分量用8位,那么一个像素共用24位表示,就说像素的深度为24)
(2)图像深度
图像深度<像素深度,接着用上面的例子,实际存储中一个像素其实是大于24位的(假设需要26位来存储),这多出来的两位可能包含了图像的一些属性。图像深度是指实际用于存储图像的灰度或色彩所需要的比特位数(24位),而此时像素深度为26位。
(3)灰度图
灰度使用黑色调来表示物体,每个灰度对象都有从0%(白色)到100%(黑色)的亮度值,灰度图就是用灰度表示的图像(说人话灰度图就是由很多个亮度不同的黑白块组成) ,灰度级越多,图像层次越逼真
(4)深度图
深度图是代表距离的图像,像素值代表图像采集器到场景中各点的距离(深度)
目前深度图的研究主要有深度估计和深度图补全,深度估计以软件为基础,通过单目或双目采集到的图像转化为深度图;深度图补全以硬件为基础,通过硬件直接拍摄出的图像存在一定的误差,需要进行一定的矫正。
2.摘要
- 背景介绍:室内语义分割一直是计算机视觉领域的一个难题
- 算法组成:在RedNet中,残差模块作为基本的积木块应用于编码器和解码器,构造了一种融合结构,提出了一种金字塔监督训练方案
- 模型评估:实验结果表明,所提出的RedNet (Resnet-50)在SUN RGBD数据集上达到了47.8%
的最新mloU
3.Pytorch实现RedNet
from torch import nn
def conv3x3(in_planes,out_planes,stride=1):
return nn.Conv2d(in_planes,out_planes,kernel_size=3,stride=stride,padding=1,bias=False)
class Bottleneck(nn.Module):
expansion=4
def __init__(self,inplanes,planes,stride=1,downsample=None):
super(Bottleneck,self).__init__()
self.conv1=nn.Conv2d(inplanes,planes,kernel_size=1,bias=False)
self.bn1=nn.BatchNorm2d(planes)
self.conv2=nn.Conv2d(planes,planes,kernel_size=3,stride=stride,padding=1,bias=False)
self.bn2=nn.BatchNorm2d(planes)
self.conv3=nn.Conv2d(planes,planes*4,kernel_size=1,bias=False)
self.bn3=nn.BatchNorm2d(planes*4)
self.relu=nn.ReLU(inplace=True)
self.downsample=downsample
self.stride=stride
def forward(self,x):
residual=x
out=self.conv1(x)
out=self.bn1(out)
out=self.relu(out)
out=self.conv2(out)
out=self.bn2(out)
out=self.relu(out)
out=self.conv3(out)
out=self.bn3(out)
if self.downsample is not None:
residual=self.downsample(x)
out+=residual
out=self.relu(out)
return out
class TransBasicBlock(nn.Module):
def __init__(self,inplanes,planes,stride=1,upsample=None):
super(TransBasicBlock,self).__init__()
self.conv1=conv3x3(inplanes,inplanes)
self.bn1=nn.BatchNorm2d(inplanes)
self.relu=nn.ReLU(inplace=True)
if upsample is not None and stride!=1:
self.conv2=nn.ConvTranspose2d(inplanes,planes,kernel_size=3,stride=stride,padding=1,output_padding=1,bias=False)
else:
self.conv2=conv3x3(inplanes,planes,stride)
self.bn2=nn.BatchNorm2d(planes)
self.upsample=upsample
self.stride=stride
def forward(self,x):
residual=x
out=self.conv1(x)
out=self.bn1(out)
out=self.relu(out)
out=self.conv2(out)
out=self.bn2(out)
if self.upsample is not None:
residual=self.upsample(x)
out+=residual
out=self.relu(out)
return out
class RedNet(nn.Module):
def __init__(self, num_classes=41):
super(RedNet, self).__init__()
block = Bottleneck
transblock = TransBasicBlock
layers = [3, 4, 6, 3]
# original resnet
self.inplanes = 64
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
# resnet for depth channel
self.inplanes = 64
self.conv1_d = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1_d = nn.BatchNorm2d(64)
self.layer1_d = self._make_layer(block, 64, layers[0])
self.layer2_d = self._make_layer(block, 128, layers[1], stride=2)
self.layer3_d = self._make_layer(block, 256, layers[2], stride=2)
self.layer4_d = self._make_layer(block, 512, layers[3], stride=2)
self.inplanes = 512
self.deconv1 = self._make_transpose(transblock, 256, 6, stride=2)
self.deconv2 = self._make_transpose(transblock, 128, 4, stride=2)
self.deconv3 = self._make_transpose(transblock, 64, 3, stride=2)
self.deconv4 = self._make_transpose(transblock, 64, 3, stride=2)
self.agant0 = self._make_agant_layer(64, 64)
self.agant1 = self._make_agant_layer(64 * 4, 64)
self.agant2 = self._make_agant_layer(128 * 4, 128)
self.agant3 = self._make_agant_layer(256 * 4, 256)
self.agant4 = self._make_agant_layer(512 * 4, 512)
# final block
self.inplanes = 64
self.final_conv = self._make_transpose(transblock, 64, 3)
self.final_deconv = nn.ConvTranspose2d(self.inplanes, num_classes, kernel_size=2,
stride=2, padding=0, bias=True)
self.out5_conv = nn.Conv2d(256, num_classes, kernel_size=1, stride=1, bias=True)
self.out4_conv = nn.Conv2d(128, num_classes, kernel_size=1, stride=1, bias=True)
self.out3_conv = nn.Conv2d(64, num_classes, kernel_size=1, stride=1, bias=True)
self.out2_conv = nn.Conv2d(64, num_classes, kernel_size=1, stride=1, bias=True)
def _make_layer(self, block, planes, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes, planes * block.expansion,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes))
return nn.Sequential(*layers)
def _make_transpose(self, block, planes, blocks, stride=1):
upsample = None
if stride != 1:
upsample = nn.Sequential(
nn.ConvTranspose2d(self.inplanes, planes,
kernel_size=2, stride=stride,
padding=0, bias=False),
nn.BatchNorm2d(planes),
)
layers = []
for i in range(1, blocks):
layers.append(block(self.inplanes, self.inplanes))
layers.append(block(self.inplanes, planes, stride, upsample))
self.inplanes = planes
return nn.Sequential(*layers)
def _make_agant_layer(self, inplanes, planes):
layers = nn.Sequential(
nn.Conv2d(inplanes, planes, kernel_size=1,
stride=1, padding=0, bias=False),
nn.BatchNorm2d(planes),
nn.ReLU(inplace=True)
)
return layers
def forward_downsample(self, rgb, depth):
x = self.conv1(rgb)
x = self.bn1(x)
x = self.relu(x)
depth = self.conv1_d(depth)
depth = self.bn1_d(depth)
depth = self.relu(depth)
fuse0 = x + depth
x = self.maxpool(fuse0)
depth = self.maxpool(depth)
x = self.layer1(x)
depth = self.layer1_d(depth)
fuse1 = x + depth
x = self.layer2(fuse1)
depth = self.layer2_d(depth)
fuse2 = x + depth
x = self.layer3(fuse2)
depth = self.layer3_d(depth)
fuse3 = x + depth
x = self.layer4(fuse3)
depth = self.layer4_d(depth)
fuse4 = x + depth
return fuse0, fuse1, fuse2, fuse3, fuse4
def forward_upsample(self, fuse0, fuse1, fuse2, fuse3, fuse4):
agent4 = self.agant4(fuse4)
x = self.deconv1(agent4)
if self.training:
out5 = self.out5_conv(x)
x = x + self.agant3(fuse3)
x = self.deconv2(x)
if self.training:
out4 = self.out4_conv(x)
x = x + self.agant2(fuse2)
x = self.deconv3(x)
if self.training:
out3 = self.out3_conv(x)
x = x + self.agant1(fuse1)
x = self.deconv4(x)
if self.training:
out2 = self.out2_conv(x)
x = x + self.agant0(fuse0)
x = self.final_conv(x)
out = self.final_deconv(x)
if self.training:
return out, out2, out3, out4, out5
return out
def forward(self, rgb, depth):
fuse = self.forward_downsample(rgb, depth)
out = self.forward_upsample(*fuse)
return out
if __name__ == "__main__":
import torch as t
rgb = t.randn(1, 3, 352, 480)
depth = t.randn(1, 1, 352, 480)
net = RedNet()
out = net(rgb, depth)
print(out[0].shape)
print(out[1].shape)
print(out[2].shape)
print(out[3].shape)
print(out[4].shape)
参考:
B站深度之眼