LoFTR： Detector-Free Local Feature Matching with Transformers

原创

已于 2024-06-08 21:43:54 修改 · 2.5k 阅读

标签

于 2024-05-26 10:35:14 首次发布

【引用格式】：Sun J, Shen Z, Wang Y, et al. LoFTR: Detector-free local feature matching with transformers[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021: 8922-8931.

【网址】https://openaccess.thecvf.com/content/CVPR2021/papers/Sun_LoFTR_Detector-Free_Local_Feature_Matching_With_Transformers_CVPR_2021_paper.pdf

【开源代码】：https://github.com/zju3dv/LoFTR

1.1 Local Feature Extraction（局部特征提取）

1.2 Local Feature Transformer (LoFTR) Module

1.3 Establishing Coarse-level Matches（建立粗级匹配）

1.4 Coarse-to-Fine Module（微调模块）

一、瓶颈问题

大多数现有的匹配放大都是有三个阶段工作：特征点检测、特征点描述和特征点匹配。然而，当在低纹理区域、重复模式、视角变化、光照变化和运动模糊等情况下难以提取可重复兴趣点，进而会导致特征点匹配失败。

二、本文贡献

提出了一种名为LoFTER的新型无检测器的局部图像特征匹配方法，使用了Transformer框架来处理图像特征，更好的获得全局上下文信息。
LoFTER首先在粗略级别建立像素级的密集匹配，然后在此基础上细化出高质量的匹配，避免了传统方法中依赖于特征检测器的局限性。
分别在室内和室外数据集上进行实验，证明了其在多个图像匹配和相机姿态估计任务重优于现有的基于检测器和无检测器的特征匹配基线方法。

三、解决方案

1 LoFTER网络框架

LoFTER首先通过对图像进行卷积下采样以及上采样等操作，获得在原始图像1/8维度处的粗粒度特征表示和1/2维度处的细粒度特征表示；然后，将粗粒度特征表示进行Transformer特征提取；将得到的特征表示图进行特征匹配，获得粗匹配；然后将细粒度特征表示重新进行Transformer操作，对粗匹配进行微调，以获得像素级别匹配。

1.1 Local Feature Extraction（局部特征提取）

局部特征提取使用了带有特征金字塔（FPN）的标准卷积框架从两个图片中提取多层级的特征。文章使用 $\tilde{F} ^{A}$ 和 $\tilde{F} ^{B}$ 表示粗略级别的特征图，大小是原始图像维度的1/8；使用 $\hat{F} ^{A}$ 和 $\hat{F} ^{B}$ 表示细级别特征图，大小是原始图像维度的1/2。

import torch.nn as nn
import torch.nn.functional as F


def conv1x1(in_planes, out_planes, stride=1):
    """1x1 convolution without padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=0, bias=False)


def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)


class BasicBlock(nn.Module):
    def __init__(self, in_planes, planes, stride=1):
        super().__init__()
        self.conv1 = conv3x3(in_planes, planes, stride)
        self.conv2 = conv3x3(planes, planes)
        self.bn1 = nn.BatchNorm2d(planes)
        self.bn2 = nn.BatchNorm2d(planes)
        self.relu = nn.ReLU(inplace=True)

        if stride == 1:
            self.downsample = None
        else:
            self.downsample = nn.Sequential(
                conv1x1(in_planes, planes, stride=stride),
                nn.BatchNorm2d(planes)
            )

    def forward(self, x):
        y = x
        y = self.relu(self.bn1(self.conv1(y)))
        y = self.bn2(self.conv2(y))

        if self.downsample is not None:
            x = self.downsample(x)

        return self.relu(x+y)


class ResNetFPN_8_2(nn.Module):
    """
    ResNet+FPN, output resolution are 1/8 and 1/2.
    Each block has 2 layers.
    """

    def __init__(self, config):
        super().__init__()
        # Config
        block = BasicBlock
        initial_dim = config['initial_dim']
        block_dims = config['block_dims']

        # Class Variable
        self.in_planes = initial_dim

        # Networks
        self.conv1 = nn.Conv2d(1, initial_dim, kernel_size=7, stride=2, padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(initial_dim)
        self.relu = nn.ReLU(inplace=True)

        self.layer1 = self._make_layer(block, block_dims[0], stride=1)  # 1/2
        self.layer2 = self._make_layer(block, block_dims[1], stride=2)  # 1/4
        self.layer3 = self._make_layer(block, block_dims[2], stride=2)  # 1/8

        # 3. FPN upsample
        self.layer3_outconv = conv1x1(block_dims[2], block_dims[2])
        self.layer2_outconv = conv1x1(block_dims[1], block_dims[2])
        self.layer2_outconv2 = nn.Sequential(
            conv3x3(block_dims[2], block_dims[2]),
            nn.BatchNorm2d(block_dims[2]),
            nn.LeakyReLU(),
            conv3x3(block_dims[2], block_dims[1]),
        )
        self.layer1_outconv = conv1x1(block_dims[0], block_dims[1])
        self.layer1_outconv2 = nn.Sequential(
            conv3x3(block_dims[1], block_dims[1]),
            nn.BatchNorm2d(block_dims[1]),
            nn.LeakyReLU(),
            conv3x3(block_dims[1], block_dims[0]),
        )

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def _make_layer(self, block, dim, stride=1):
        layer1 = block(self.in_planes, dim, stride=stride)
        layer2 = block(dim, dim, stride=1)
        layers = (layer1, layer2)

        self.in_planes = dim
        return nn.Sequential(*layers)

    def forward(self, x):
        # ResNet Backbone
        x0 = self.relu(self.bn1(self.conv1(x)))
        x1 = self.layer1(x0)  # 1/2
        x2 = self.layer2(x1)  # 1/4
        x3 = self.layer3(x2)  # 1/8

        # FPN
        x3_out = self.layer3_outconv(x3)

        x3_out_2x = F.interpolate(x3_out, scale_factor=2., mode='bilinear', align_corners=True)
        x2_out = self.layer2_outconv(x2)
        x2_out = self.layer2_outconv2(x2_out+x3_out_2x)

        x2_out_2x = F.interpolate(x2_out, scale_factor=2., mode='bilinear', align_corners=True)
        x1_out = self.layer1_outconv(x1)
        x1_out = self.layer1_outconv2(x1_out+x2_out_2x)

        return [x3_out, x1_out]


class ResNetFPN_16_4(nn.Module):
    """
    ResNet+FPN, output resolution are 1/16 and 1/4.
    Each block has 2 layers.
    """

    def __init__(self, config):
        super().__init__()
        # Config
        block = BasicBlock
        initial_dim = config['initial_dim']
        block_dims = config['block_dims']

        # Class Variable
        self.in_planes = initial_dim

        # Networks
        self.conv1 = nn.Conv2d(1, initial_dim, kernel_size=7, stride=2, padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(initial_dim)
        self.relu = nn.ReLU(inplace=True)

        self.layer1 = self._make_layer(block, block_dims[0], stride=1)  # 1/2
        self.layer2 = self._make_layer(block, block_dims[1], stride=2)  # 1/4
        self.layer3 = self._make_layer(block, block_dims[2], stride=2)  # 1/8
        self.layer4 = self._make_layer(block, block_dims[3], stride=2)  # 1/16

        # 3. FPN upsample
        self.layer4_outconv = conv1x1(block_dims[3], block_dims[3])
        self.layer3_outconv = conv1x1(block_dims[2], block_dims[3])
        self.layer3_outconv2 = nn.Sequential(
            conv3x3(block_dims[3], block_dims[3]),
            nn.BatchNorm2d(block_dims[3]),
            nn.LeakyReLU(),
            conv3x3(block_dims[3], block_dims[2]),
        )

        self.layer2_outconv = conv1x1(block_dims[1], block_dims[2])
        self.layer2_outconv2 = nn.Sequential(
            conv3x3(block_dims[2], block_dims[2]),
            nn.BatchNorm2d(block_dims[2]),
            nn.LeakyReLU(),
            conv3x3(block_dims[2], block_dims[1]),
        )

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def _make_layer(self, block, dim, stride=1):
        layer1 = block(self.in_planes, dim, stride=stride)
        layer2 = block(dim, dim, stride=1)
        layers = (layer1, layer2)

        self.in_planes = dim
        return nn.Sequential(*layers)

    def forward(self, x):
        # ResNet Backbone
        x0 = self.relu(self.bn1(self.conv1(x)))
        x1 = self.layer1(x0)  # 1/2
        x2 = self.layer2(x1)  # 1/4
        x3 = self.layer3(x2)  # 1/8
        x4 = self.layer4(x3)  # 1/16

        # FPN
        x4_out = self.layer4_outconv(x4)

        x4_out_2x = F.interpolate(x4_out, scale_factor=2., mode='bilinear', align_corners=True)
        x3_out = self.layer3_outconv(x3)
        x3_out = self.layer3_outconv2(x3_out+x4_out_2x)

        x3_out_2x = F.interpolate(x3_out, scale_factor=2., mode='bilinear', align_corners=True)
        x2_out = self.layer2_outconv(x2)
        x2_out = self.layer2_outconv2(x2_out+x3_out_2x)

        return [x4_out, x2_out]

1.2 Local Feature Transformer (LoFTR) Module

输入特征：LoFTER模块接收来自局部特征提取阶段的粗略级别特征图 $\tilde{F} ^{A}$ 和

最低0.47元/天解锁文章