Android Q - 从人像模式照片中获取深度图

Question

我正在尝试构建一个示例 Android 应用程序来提取使用谷歌相机应用程序拍摄的人像模式照片的深度图。我知道它是沿着模糊的照片保存的。

我阅读了来自 Google 的动态深度格式文档:https://developer.android.com/training/camera2/Dynamic-depth-v1.0.pdf 这是非常新的，我找不到与此主题相关的任何资源或如何管理 android 肖像的深度图的提取。

我使用元数据提取器库以编程方式读取文件的元数据，尤其是 XMP 部分，因为它是描述深度信息的地方。我用 Kotlin 编写了一个示例代码来尝试提取深度图:

val inputStream = contentResolver.openInputStream(imageUri)

inputStream?.let { stream ->
    val metadata = JpegMetadataReader.readMetadata(stream)
    val directories = metadata.directories
    val xmpDirectories = metadata.getDirectoriesOfType(XmpDirectory::class.java)

    for (xmpDirectory in xmpDirectories) {
        val xmpMeta = xmpDirectory.xmpMeta
        val itr = xmpMeta.iterator()

        while (itr.hasNext()) {
            val propertyInfo = itr.next() as XMPPropertyInfo

            println(propertyInfo.path + " :: " + propertyInfo.value)
        }
    }

    stream.close()
}

结果输出:

    xmpNote:HasExtendedXMP :: 5c970bbab778024b23c5a8269325455c
    null :: null
    GCreations:CameraBurstID :: 9e286063-a919-4a74-96ee-d7e02d2a17d2
    null :: null
    GCamera:BurstID :: 9e286063-a919-4a74-96ee-d7e02d2a17d2
    GCamera:BurstPrimary :: 1
    GCamera:SpecialTypeID :: 
    GCamera:SpecialTypeID[1] :: com.google.android.apps.camera.gallery.specialtype.SpecialType-PORTRAIT
    null :: null
    Device:Container :: 
    Device:Container/Container:Directory :: 
    Device:Container/Container:Directory[1] :: 
    Device:Container/Container:Directory[1]/Item:Mime :: image/jpeg
    Device:Container/Container:Directory[1]/Item:Length :: 0
    Device:Container/Container:Directory[1]/Item:DataURI :: primary_image
    Device:Container/Container:Directory[1]/rdf:type :: http://ns.google.com/photos/dd/1.0/container/:Item
    Device:Container/Container:Directory[2] :: 
    Device:Container/Container:Directory[2]/Item:Mime :: image/jpeg
    Device:Container/Container:Directory[2]/Item:Length :: 1499039
    Device:Container/Container:Directory[2]/Item:DataURI :: android/original_image
    Device:Container/Container:Directory[2]/rdf:type :: http://ns.google.com/photos/dd/1.0/container/:Item
    Device:Container/Container:Directory[3] :: 
    Device:Container/Container:Directory[3]/Item:Mime :: image/jpeg
    Device:Container/Container:Directory[3]/Item:Length :: 316885
    Device:Container/Container:Directory[3]/Item:DataURI :: android/depthmap
    Device:Container/Container:Directory[3]/rdf:type :: http://ns.google.com/photos/dd/1.0/container/:Item
    Device:Container/Container:Directory[4] :: 
    Device:Container/Container:Directory[4]/Item:Mime :: image/jpeg
    Device:Container/Container:Directory[4]/Item:Length :: 65189
    Device:Container/Container:Directory[4]/Item:DataURI :: android/confidencemap
    Device:Container/Container:Directory[4]/rdf:type :: http://ns.google.com/photos/dd/1.0/container/:Item
    Device:Profiles :: 
    Device:Profiles[1] :: 
    Device:Profiles[1]/Profile:Type :: DepthPhoto
    Device:Profiles[1]/Profile:CameraIndices :: 
    Device:Profiles[1]/Profile:CameraIndices[1] :: 0
    Device:Profiles[1]/rdf:type :: http://ns.google.com/photos/dd/1.0/device/:Profile
    Device:Cameras :: 
    Device:Cameras[1] :: 
    Device:Cameras[1]/Camera:Trait :: Physical
    Device:Cameras[1]/Camera:Image :: 
    Device:Cameras[1]/Camera:Image/Image:ItemSemantic :: Original
    Device:Cameras[1]/Camera:Image/Image:ItemURI :: android/original_image
    Device:Cameras[1]/Camera:DepthMap :: 
    Device:Cameras[1]/Camera:DepthMap/DepthMap:ItemSemantic :: Depth
    Device:Cameras[1]/Camera:DepthMap/DepthMap:Format :: RangeInverse
    Device:Cameras[1]/Camera:DepthMap/DepthMap:Units :: Diopters
    Device:Cameras[1]/Camera:DepthMap/DepthMap:Near :: 0.302570
    Device:Cameras[1]/Camera:DepthMap/DepthMap:Far :: 1.754560
    Device:Cameras[1]/Camera:DepthMap/DepthMap:DepthURI :: android/depthmap
    Device:Cameras[1]/Camera:DepthMap/DepthMap:MeasureType :: OpticalAxis
    Device:Cameras[1]/Camera:DepthMap/DepthMap:ConfidenceURI :: android/confidencemap
    Device:Cameras[1]/Camera:DepthMap/DepthMap:FocalTableEntryCount :: 256
    Device:Cameras[1]/Camera:DepthMap/DepthMap:FocalTable :: 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
    Device:Cameras[1]/Camera:ImagingModel :: 
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:FocalLengthX :: 3187.589355
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:FocalLengthY :: 3187.589355
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:ImageWidth :: 4032
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:ImageHeight :: 3024
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:PrincipalPointX :: 2000.483154
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:PrincipalPointY :: 1541.417236
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:Skew :: 0.000000
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:PixelAspectRatio :: 1.000000
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:DistortionCount :: 4
    Device:Cameras[1]/Camera:ImagingModel/ImagingModel:Distortion :: AACAPwAAAABBQ7w9AAAAAEgGf74AAAAA5SpFPgAAAAA
    Device:Cameras[1]/rdf:type :: http://ns.google.com/photos/dd/1.0/device/:Camera

根据 Google 的文档，深度图图像被序列化为 base64 字符串 XMP 属性。但我不知道如何提取它以生成基于此深度数据的新图像。我认为我几乎解决了我的问题，但我错过了一些关于 Adob​​e XMP 标准的确认。

我发现了一个叫做“sidecar Xmp 文件”的东西，也许我试图找到的深度图就在其中。

我上传到https://www.photopea.com上，成功看到深度图嵌入了照片

Answer 1

像素手机人像模式照片由4个JFIF结构拼接https://en.wikipedia.org/wiki/JPEG_File_Interchange_Format .每个 JFIF 结构都是一个 jpeg 图像。

JFIF 结构以标记 0xFFD8 开始，以标记 0xFFD9 结束。因此，我们可以将一个人像模式图像拆分为 4 个 jpeg 文件。

以下 python 代码打印标记位置并将 PXL_20210107_114027740.PORTRAIT.jpg 拆分为，

1. pxl_out_0.jpg: 显示图像
2. pxl_out_1.jpg:原图
3. pxl_out_2.jpg:256级灰度的深度图
4. pxl_out_3.jpg:填充了 255 的虚拟图像

with open('PXL_20210107_114027740.PORTRAIT.jpg', mode='rb') as infile:
    buffer = infile.read()

bufferlen = len(buffer)
pos = 0
pos_d8 = 0
n = 0
i = 0
while i < bufferlen:
    if buffer[i] == 0xff:
        pos = i
        i += 1
        if buffer[i] == 0xd8:
            print('ffd8: {0}'.format(pos))
            pos_d8 = pos
        elif buffer[i] == 0xd9:
            print('ffd9: {0} len: {1}'.format(pos, pos - pos_d8 + 2))
            with open('pxl_out_{0}.jpg'.format(n), mode='wb') as outfile:
                n += 1
                outfile.write(buffer[pos_d8: pos + 2])
    i += 1