NB: The usual blog disclaimer for this site applies to posts around HoloLens. I am not on the HoloLens team. I have no details on HoloLens other than what is on the public web and so what I post here is just from my own experience experimenting with pieces that are publicly available and you should always check out the official developer site for the product documentation.
For quite a while, I have wanted to put together a small sample that ran on HoloLens which did facial tracking but which also identified in X,Y,Z space where the tracked faces were located.
I should start by saying that there is an official sample which does exactly this and it’s over here on github;
and so if you’re looking for a sample to learn from which actually works 
then feel free to stop reading at this point.
I spent a little time looking at this sample and when I dug into it, I noticed that it takes a (reasonable) approach to estimating the Z co-ordinate of the face that has been detected in an RGB frame by assuming that a face at a certain distance (1m) will have a certain width and using that as a basis for working out how far away the face is.
That seems to work pretty well but I’d long since been drawn by the idea that the HoloLens could give me a more accurate measurement of how far away a face was and I wanted to experiment with that.
The rest of this post is my rough notes around some experiments that don’t quite get me to the point of making that work but I wrote things down as I went along thinking that the notes might be useful to me in the future and (maybe) to someone else.
I’d urge you to apply a large pinch of salt to what you see here because I’m very much experimenting and definitely don’t have things figured out fully.
That sample is great though because it brings together a number of the pieces that I’d expect to use in my own experiments such as;
- It uses the media capture APIs to get access to a web camera and read frames from it – classes like MediaFrameSource and MediaFrameReader
- It uses the CameraIntrinsics class in order to ‘unproject’ from the pixel X,Y co-ordinates of a web-camera-captured-image back into a co-ordinate system of the web camera itself
- It uses the SpatialCoordinateSystem class to transform from a co-ordinate system of a web-camera back into the world co-ordinate system of the application itself
- It uses the FaceTracker class in order to identify faces in images taken from the web-camera (there’s no need to call out to the cloud just to detect the bounding rectangles of a face)
The sample also takes an approach that is similar to the one that I wanted to follow in that it tries to do ‘real time’ processing on video frames taken from a stream coming off the camera rather than e.g. taking a single photo image or recording frames into a file or something in order to do some type of ‘one-shot’ or ‘offline’ processing.
In many ways, then, this sample is what I want but I have a few differences in what I want to do;
- This sample estimates the Z coordinate of the face whereas I’d like to ask the device to give it to me as accurately as I can get.
- This sample is written in C++ using DirectX whereas I wanted to do something in C# and Unity as C++ is becoming for me a language that I can better read than write as I do so little of it these days.
Getting to the Z-coordinate feels like it’s the main challenge here and I can think of at least a couple of ways of doing it. One might be to make use of the spatial mesh that the HoloLens makes available and use some kind of ‘ray cast’ from the camera along the user’s gaze vector to calculate the length of the ray that hits the mesh and use that as a depth value. For a long time, this sort of idea has been discussed in the HoloLens developer forums;
but it initially felt to me like it might not be quite right for tracking people as they moved around in a room and so rather than going in that direction, I wanted to experiment with the new ‘research mode’ which came to HoloLens in the Redstone 4 release and which gives me direct access to a stream of depth images. It seemed like it might be ‘interesting’ to see what it’s like to directly use the depth frames from the camera in order to calculate a Z-coordinate for a pixel in an image taken from the web (RGB) camera.
I first encountered ‘Research Mode’ in this blog post;
Experimenting with Research Mode and Sensor Streams on HoloLens Redstone 4 Preview
and there are a tonne of caveats around using ‘Research Mode’ and so please read that post and the official docs and make sure that you understand the caveats before you do anything with it on your own devices.
Note also that if you want something more definitive (and correct!) around research mode then some new samples were published while I was experimenting around for this post and so make sure that you check those out here;
https://github.com/Microsoft/HoloLensForCV
With that said, what is it that I wanted to achieve in my sample?
- Read video frames from the RGB camera at some frequency – this is achievable using the UWP media capture APIs.
- Identify faces within those video frames – achievable using the UWP face detection APIs which will give me bounding boxes (X,Y, width, height) in pixel co-ordinates within the captured image. I could perhaps simplify this into a single X,Y point per face by taking the centre of the bounding box.
- Read depth frames from the depth camera at some frequency – this is achievable using the UWP media capture APIs on a device running in ‘research mode’.
- Somehow correlate a depth frame with a video frame.
- Use the X,Y co-ordinate of the detected face( s ) in a video frame to index into a frame from the depth camera to determine the depth location at that point.
- Transform the X,Y co-ordinate from the video frame and the Z co-ordinate from the depth frame back into world space in order to display something (e.g. a box or a 3D face) at the location that the face was detected.
This leaves me with a sort of 50:50 ratio between things that I know can be done and things that I haven’t done before and so I set about a few experiments to test a few things out.
Before doing that though I spent a long time reading a couple of articles in the docs;
and it’s worth saying that the LocatableCamera support in Unity is very close to what I want but, unfortunately, the Video Capture side of those APIs only record video to a file rather than giving me a stream of video frames which I can then process using something like the Face detection APIs and so I don’t think it’s really of use to me.
Other people have noticed this here;
HoloLensCameraStream for Unity
although at the time of writing I haven’t dug too far into that project as I encountered it when I’d nearly completed this post.
With all that said, I switched on ‘research mode’ on my HoloLens and tried a few initial experiments that I tried out to see how things looked…
Experiment 1 – Co-ordinating Frames from Multiple Sources
If I want to read video frames and depth frames at the same time then the UWP has the notion of a MultiSourceMediaFrameReader which can do just that for me firing an event when a frame is available from each of the requested media sources and saving me a bunch of work in trying to piece those frames together myself.
I was curious – could I get a frame from the RGB video camera and a frame from the depth camera at the same time pre-correlated by the system for me?
In looking a little deeper, I’m not sure that I can. As far as I can tell, the only way to get hold of a MultiSourceMediaFrameReader is to call MediaCapture.CreateMultiSourceFrameReaderAsync and that requires a MediaCapture which can be initialised for one and only MediaFrameSourceGroup and as the docs say;
“The MediaFrameSourceGroup object represents a set of media frame sources that can be used simultaneously”
and so if I run this little piece of code on my HoloLens;
var colourGroup = await MediaFrameSourceGroup.FindAllAsync();
foreach (var group in colourGroup)
{
Debug.WriteLine($"Group {group.DisplayName}");
foreach (var source in group.SourceInfos)
{
Debug.WriteLine($"\tSource {source.MediaStreamType}, {source.SourceKind}, {source.Id}");
foreach (var profile in source.VideoProfileMediaDescription)
{
Debug.WriteLine($"\t\tProfile {profile.Width}x{profile.Height}@{profile.FrameRate:N0}");
}
}
}
then the output I get is as below;
Group Sensor Streaming
Source VideoRecord, Depth, Source#0@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 448×450@15
Source VideoRecord, Infrared, Source#1@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 448×450@3
Source VideoRecord, Depth, Source#2@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 448×450@15
Source VideoRecord, Infrared, Source#3@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 448×450@3
Source VideoRecord, Color, Source#4@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 160×480@30
Source VideoRecord, Color, Source#5@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 160×480@30
Source VideoRecord, Color, Source#6@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 160×480@30
Source VideoRecord, Color, Source#7@\\?\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}
Profile 160×480@30
Group MN34150
Source VideoPreview, Color, Source#0@\\?\DISPLAY#INT22B8#4&27b432bd&0&UID139960#{e5323777-f976-4f5b-9b55-b94699c46e44}\{cdd6871a-56ca-4386-bae7-d24b564378a9}
Profile 1280×720@30
Profile 896×504@30
Profile 1344×756@30
Profile 1408×792@30
Profile 1280×720@24
Profile 896×504@24
Profile 1344×756@24
Profile 1408×792@24
Profile 1280×720@20
Profile 896×504@20
Profile 1344×756@20
Profile 1408×792@20
Profile 1280×720@15
Profile 896×504@15
Profile 1344×756@15
Profile 1408×792@15
Profile 1280×720@5
Profile 896×504@5
Profile 1344×756@5
Profile 1408×792@5
Source VideoRecord, Color, Source#1@\\?\DISPLAY#INT22B8#4&27b432bd&0&UID139960#{e5323777-f976-4f5b-9b55-b94699c46e44}\{cdd6871a-56ca-4386-bae7-d24b564378a9}
Profile 1280×720@30
Profile 896×504@30
Profile 1344×756@30
Profile 1408×792@30
Profile 1280×720@24
Profile 896×504@24
Profile 1344×756@24
Profile 1408×792@24
Profile 1280×720@20
Profile 896×504@20
Profile 1344×756@20
Profile 1408×792@20
Profile 1280×720@15
Profile 896×504@15
Profile 1344×756@15
Profile 1408×792@15
Profile 1280×720@5
Profile 896×504@5
Profile 1344×756@5
Profile 1408×792@5
Source Photo, Image, Source#2@\\?\DISPLAY#INT22B8#4&27b432bd&0&UID139960#{e5323777-f976-4f5b-9b55-b94699c46e44}\{cdd6871a-56ca-4386-bae7-d24b564378a9}
Profile 1280×720@30
Profile 1280×720@0
Profile 896×504@30
Profile 896×504@0
Profile 1344×756@30
Profile 1344×756@0
Profile 1408×792@30
Profile 1408×792@0
Profile 2048×1152@30
Profile 2048×1152@0
I’m not sure why some profiles seem to come up with a zero frame-rate, I perhaps need to look at that but I read this as essentially telling me that I have 2 MediaFrameSourceGroups here and my RGB stream is in one and my depth streams are in another and so I don’t think that I can use a single MediaCapture and a single MultiSourceMediaFrameReader to map between them.
I think that leaves me with a couple of options;
- I could try and use a multi source frame reader across Depth + InfraRed and see whether I can do facial detection on the InfraRed images?
- I could avoid multi source frame readers, use separate readers and take some approach to trying to correlate depth and RGB images myself.
The other thing that surprised me here is that the frame rates of those depth streams (both reported at 15fps) don’t seem to line up with the Research Mode docs – I’ll come back to this.
This was a useful experiment and my inclination is to go with the second approach – have multiple frame readers and try to link up frames as best that I can.
Experiment 2 – Camera Intrinsics, Coordinate Systems
When you receive frames from a media source, they are delivered in the shape of a MediaFrameReference which contains metadata such as timings, durations, formats and a CoordinateSystem and then the VideoMediaFrame itself.
That frame then provides access to (e.g.) the SoftwareBitmap (if it’s been requested by choosing a Cpu preference) and if the frame is a depth frame then it also offers up details on that depth data via the DepthMediaFrame property.
If then add a little code into my example to try and create MediaCapture and MediaFrameSource instances for me as below;
async Task<(MediaCapture capture, MediaFrameSource source)> GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind sourceKind,
int width,
int height,
int frameRate)
{
MediaCapture mediaCapture = null;
MediaFrameSource frameSource = null;
var allSources = await MediaFrameSourceGroup.FindAllAsync();
// Ignore frame rate here on the description as both depth streams seem to tell me they are
// 30fps whereas I don't think they are (from the docs) so I leave that to query later on.
var sourceInfo =
allSources.SelectMany(group => group.SourceInfos)
.FirstOrDefault(
si =>
(si.MediaStreamType == MediaStreamType.VideoRecord) &&
(si.SourceKind == sourceKind) &&
(si.VideoProfileMediaDescription.Any(
desc =>
desc.Width == width &&
desc.Height == height &&
desc.FrameRate == frameRate)));
if (sourceInfo != null)
{
var sourceGroup = sourceInfo.SourceGroup;
mediaCapture = new MediaCapture();
await mediaCapture.InitializeAsync(
new MediaCaptureInitializationSettings()
{
// I want software bitmaps
MemoryPreference = MediaCaptureMemoryPreference.Cpu,
SourceGroup = sourceGroup,
StreamingCaptureMode = StreamingCaptureMode.Video
}
);
frameSource = mediaCapture.FrameSources[sourceInfo.Id];
var selectedFormat = frameSource.SupportedFormats.First(
format => format.VideoFormat.Width == width && format.VideoFormat.Height == height &&
format.FrameRate.Numerator / format.FrameRate.Denominator == frameRate);
await frameSource.SetFormatAsync(selectedFormat);
}
return (mediaCapture, frameSource);
}
then I can open up both an RGB frame reader and a depth frame reader and have a bit of a look at what’s present there…
var rgbMedia = await this.GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind.Color, 1280, 720, 30);
var rgbReader = await rgbMedia.capture.CreateFrameReaderAsync(rgbMedia.source);
rgbReader.AcquisitionMode = MediaFrameReaderAcquisitionMode.Realtime;
var depthMedia = await this.GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind.Depth, 448, 450, 15);
var depthReader = await depthMedia.capture.CreateFrameReaderAsync(depthMedia.source);
depthReader.AcquisitionMode = MediaFrameReaderAcquisitionMode.Realtime;
TypedEventHandler<MediaFrameReader, MediaFrameArrivedEventArgs> handler =
(s,e) =>
{
using (var frame = s.TryAcquireLatestFrame())
{
if (frame != null)
{
Debug.WriteLine($"Frame of type {frame.SourceKind}");
Debug.WriteLine($"Intrinsics present? {frame.VideoMediaFrame.CameraIntrinsics != null}");
Debug.WriteLine($"Coordinate system present? {frame.CoordinateSystem != null}");
}
}
};
rgbReader.FrameArrived += handler;
depthReader.FrameArrived += handler;
await rgbReader.StartAsync();
await depthReader.StartAsync();
// Wait forever then dispose...
await Task.Delay(-1);
rgbReader.Dispose();
depthReader.Dispose();
rgbMedia.capture.Dispose();
depthMedia.capture.Dispose();
and the output that I get is a little disappointing…
Frame of type Color
Intrinsics present? False
Coordinate system present? False
Frame of type Depth
Intrinsics present? False
Coordinate system present? False
so I don’t seem to get CameraIntrinsics or a CoordinateSystem on either of these frame types and I was thinking that I’d probably need both of these things in order to be able to transform from an X,Y pixel co-ordinate to world space.
I was especially hoping that the CameraIntrinsics might enable me to use this API on the DepthMediaFrame;
DepthMediaFrame.TryCreateCoordinateMapper
which sounds like it might be exactly what I need to transform points and I’ve seen this used in samples.
That lack of CameraIntrinsics seems to be picked up in this forum post;
and I wonder whether this might be a bug or maybe I’m missing some flag to switch it on but I haven’t figured that out at the time of writing.
I did also attempt to get the CameraIntrinsics by reaching into the MediaFrameSource and using the TryGetCameraIntrinsics method but I found that this seemed to return NULL for all the combinations of parameters that I passed to it.
What to do? Well, that locatable camera article suggested that there may be more than one way to go about this, specifically if I have these 3 GUIDs;
static readonly Guid MFSampleExtension_Spatial_CameraCoordinateSystem = new Guid("9D13C82F-2199-4E67-91CD-D1A4181F2534");
static readonly Guid MFSampleExtension_Spatial_CameraProjectionTransform = new Guid("47F9FCB5-2A02-4F26-A477-792FDF95886A");
static readonly Guid MFSampleExtension_Spatial_CameraViewTransform = new Guid("4E251FA4-830F-4770-859A-4B8D99AA809B");
then I can index into the properties that are part of the MediaFrameReference and if I change my innermost if statement to be;
if (frame != null)
{
Debug.WriteLine($"Frame of type {frame.SourceKind}");
SpatialCoordinateSystem coordinateSystem = null;
byte[] viewTransform = null;
byte[] projectionTransform = null;
object value;
if (frame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraCoordinateSystem, out value))
{
coordinateSystem = value as SpatialCoordinateSystem;
}
if (frame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraProjectionTransform, out value))
{
projectionTransform = value as byte[];
}
if (frame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraViewTransform, out value))
{
viewTransform = value as byte[];
}
Debug.WriteLine($"Coordinate system present? {coordinateSystem != null}");
Debug.WriteLine($"View transform present? {viewTransform != null}");
Debug.WriteLine($"Projection transform present? {projectionTransform != null}");
}
then I get output that indicates that I can get hold of the SpatialCoordinateSystem and the View Transform and Projection Transform for the RGB camera. I get nothing back for the depth camera.
So, that’s been a useful experiment – it tells me that I might be able to transform from an X,Y pixel co-ordinate back to world space although I need to be able to translate those byte[] arrays back into matrices.
I’m not sure quite how to do that but I wrote;
static Matrix4x4 ByteArrayToMatrix(byte[] bits)
{
Matrix4x4 matrix = Matrix4x4.Identity;
var handle = GCHandle.Alloc(bits, GCHandleType.Pinned);
matrix = Marshal.PtrToStructure<Matrix4x4>(handle.AddrOfPinnedObject());
handle.Free();
return (matrix);
}
and maybe that will do it for me if my assumption about how those matrices have been packed as byte[] is right?
Experiment 3 – Getting Depth Values
I wondered what it looked like to get depth values from the depth sensor and so I reworked the code above a little to bring in the infamous IMemoryBufferByteAccess (meaning that I have to compile with unsafe code);
[ComImport]
[Guid("5B0D3235-4DBA-4D44-865E-8F1D0E4FD04D")]
[InterfaceType(ComInterfaceType.InterfaceIsIUnknown)]
unsafe interface IMemoryBufferByteAccess
{
void GetBuffer(out byte* buffer, out uint capacity);
}
and then reworked my frame handling code so as to look as below;
var depthMedia = await this.GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind.Depth, 448, 450, 15);
var depthReader = await depthMedia.capture.CreateFrameReaderAsync(depthMedia.source);
depthReader.AcquisitionMode = MediaFrameReaderAcquisitionMode.Realtime;
TypedEventHandler<MediaFrameReader, MediaFrameArrivedEventArgs> handler =
(s,e) =>
{
using (var frame = s.TryAcquireLatestFrame())
{
if (frame != null)
{
var centrePoint = new Point(
frame.Format.VideoFormat.Width / 2,
frame.Format.VideoFormat.Height / 2);
using (var bitmap = frame.VideoMediaFrame.SoftwareBitmap)
using (var buffer = bitmap.LockBuffer(BitmapBufferAccessMode.Read))
using (var reference = buffer.CreateReference())
{
var description = buffer.GetPlaneDescription(0);
var bytesPerPixel = description.Stride / description.Width;
Debug.Assert(bytesPerPixel == Marshal.SizeOf<UInt16>());
int offset =
(description.StartIndex + description.Stride * (int)centrePoint.Y) +
((int)centrePoint.X * bytesPerPixel);
UInt16 depthValue = 0;
unsafe
{
byte* pBits;
uint size;
var byteAccess = reference as IMemoryBufferByteAccess;
byteAccess.GetBuffer(out pBits, out size);
depthValue = *(UInt16*)(pBits + offset);
}
Debug.WriteLine($"Depth in centre is {depthValue}");
}
}
}
};
depthReader.FrameArrived += handler;
await depthReader.StartAsync();
// Wait forever then dispose...
await Task.Delay(-1);
depthReader.Dispose();
depthMedia.capture.Dispose();
and so my hope is to trace out the depth value that is obtained from the centre point of the depth frame itself. I ran this and saw this type of output;
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 4.09m
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 4.09m
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 0.43m
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 0.577m
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 0.579m
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 0.583m
and the code seemed to be ‘working’ except that I noticed that when I pointed the device at more distant objects (perhaps > 0.7m) it seemed to be coming back (consistently) with a 4.09m (4090 or FFA) value which felt like some kind of ‘out of range’ regardless of the fact that the maximum reliable depth is being reported as 65m (which seems a little unlikely! ).
I can only assume that this is the ‘near’ sensor and its ID seems to be;
“Source#0@\\\\?\\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}”
I know that the HoloLens has 2 depth streams described in the documents here as;
“Two versions of the depth camera data – one for high-frequency (30 FPS) near-depth sensing, commonly used in hand tracking, and the other for lower-frequency (1 FPS) far-depth sensing, currently used by Spatial Mapping”
Now, my device seems to report two depth streams of the same dimensions (448 x 450) and of the same frame rate (15fps) so that doesn’t seem to line up with the docs.
Putting that to one side, my code had been written to simply select whichever sensor matching my search criteria came First() (in the LINQ sense) and to ignore any others.
I switched the code to select the Last() and saw values;
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 1.562m
Max Depth in Metres 65.535m
Min Depth in Metres 0.001m
Depth Format is D16
Depth in centre is 4.09m
and so it now seemed that the device was returning 4.09m (4090) for surfaces that were nearer to it (perhaps < 0.7m away) while it was correctly reporting the more distant surfaces.
I can only assume that this device is the ‘far sensor’ and its ID seems to be;
“Source#2@\\\\?\\Root#SensorStreamingMiniDriver#0000#{e5323777-f976-4f5b-9b55-b94699c46e44}\\{b27e3887-ad10-4a4e-bfb8-d6765add0e38}”
I guess that if you wanted to work reliably, you might have to take streams from both depth cameras and use whichever gave you a reliable result but for my purposes I’m going to go with the ‘far’ sensor rather than the ‘near’ sensor.
One other thing that I’d flag here – I didn’t seem to see 15fps from the depth stream, the rate seemed more like 1fps which ties up with the Research Mode docs for the ‘long range’ depth sensor so maybe the API reporting 15fps isn’t right here.
Experiment 5 – Tracking Faces
Ironically, the detection of faces feels like “the easy part” of what I’m trying to do here and it’s purely because the UWP already has APIs which track faces for me so it’s not such a big deal to make use of it.
I can take my code which is already getting hold of video frames (with SoftwareBitmaps) and just try and feed them through a FaceDetector or FaceTracker and it will give me back lists of bounding boxes of the faces that it detects.
The only potential ‘fly in the ointment’ here is that the detection requires bitmaps in specific formats and there’s an API for querying which formats are supported but that means that I need to try and either;
- Ensure that I ask the media capture APIs to hand me back bitmaps in one of the formats that is supported by the face detection APIs.
or
- Accept that the media capture APIs might not be able to do that and so gracefully fallback and accept some other format which I then convert on a frame-by-frame basis to one of the ones supported by the face detection APIs.
The second option is the ‘right’ way to do things but it means writing a bit more code and so I left the conversion for ‘another day’ and modified my GetMediaCaptureForDescriptionAsync method (not repeated here) to take an additional (optional) parameter which lets me specify that I want to narrow down my search for a media source to include specifying the set of bitmap formats that I’m prepared to accept;
var supportedFormats = FaceTracker.GetSupportedBitmapPixelFormats().Select(
format => format.ToString().ToLower()).ToArray();
var tracker = await FaceTracker.CreateAsync();
// We are assuming (!) that we can get frames in a format compatible with the
// FaceTracker.
var rgbMedia = await this.GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind.Color, 1280, 720, 30,
supportedFormats);
var rgbReader = await rgbMedia.capture.CreateFrameReaderAsync(rgbMedia.source);
rgbReader.AcquisitionMode = MediaFrameReaderAcquisitionMode.Realtime;
TypedEventHandler<MediaFrameReader, MediaFrameArrivedEventArgs> handler =
async (s, e) =>
{
using (var frame = s.TryAcquireLatestFrame())
{
if (frame != null)
{
using (var videoFrame = frame.VideoMediaFrame.GetVideoFrame())
{
var faces = await tracker.ProcessNextFrameAsync(videoFrame);
foreach (var face in faces)
{
Debug.WriteLine($"Face found at {face.FaceBox.X}, {face.FaceBox.Y}");
}
}
}
}
};
rgbReader.FrameArrived += handler;
await rgbReader.StartAsync();
// Wait forever then dispose...
await Task.Delay(-1);
rgbReader.Dispose();
rgbMedia.capture.Dispose();
and that seemed to work quite nicely – I’m getting video frames from the RGB camera and finding faces within them.
It’s worth possibly saying that in doing this I came (once again) across the place where MediaFrameFormat.Subtype which contains names for the Subtypes taken from this doc page and matching that up to BitmapPixelFormat feels like a very imprecise science and the docs even have a warning around these subtypes;
“The string values returned by the MediaEncodingSubtypes properties may not use the same letter casing as AudioEncodingProperties.Subtype, VideoEncodingProperties.Subtype, ContainerEncodingProperties.Subtype, and ImageEncodingProperties.Subtype. For this reason, if you compare the values, you should use a case-insensitive comparison or use hardcoded strings that match the casing returned by the encoding properties.”
Experiment 6 – Turning X,Y co-ordinates in Images into X,Y,Z co-ordinates in (Unity) World Space
This is the part where I get stuck 
I did mention before that I read this article about the locatable camera quite a few times because it seemed to be very relevant to what I’m trying to do;
Locatable Camera
and I especially focused on the section entitled;
Pixel to Application-specified Coordinate System
and its promise of being able to convert from pixel co-ordinates back to world co-ordinates using the camera projection matrix which (I think) I have available to me based on Experiment 2 above.
To experiment with this, I wondered whether I could take the 4 pixel points that represent the camera’s bounding box, project them back from the image into camera space and then into world space and see what they ‘looked like’ in world space by drawing them in Unity at some specified distance.
In doing that, there are perhaps a few things that I’d comment on which may be right/wrong.
- Getting hold of the SpatialCoordinateSystem that Unity sets up for my holographic app seems to be a matter of calling WorldManager.GetNativeISpatialCoordinateSystemPtr and using Marshal.* methods to get a handle onto the underlying object although I’m unclear whether it’s ok to just hold on to this object indefinitely or not.
- In transforming back from an X,Y image co-ordinate to a X,Y,Z co-ordinate in world space my approach (following the locatable camera article again) has been to;
- Translate the X,Y coordinate from the 0-1280, 0-720 range into a –1 to 1, –1 to 1 range.
- Unproject the vector using the projection transform at a unit distance
- Multiply the unprojected vector by the inverse of the view transform
- Multiply that value by the camera to world transform obtained asking the SpatialCoordinateSystem of the RGB frame to provide a transform to the SpatialCoordinateSystem that Unity has set up for the app
- Multiplying the Z co-ordinate by –1.0f as Unity is a left-handed coordinate system and the holographic UWP APIs are right handed.
I’m not at all sure that I have this right and I was especially unsure around whether the SpatialCoordinateSystem of a frame would change as the device moved around and/or whether the view transform would change. I used the debugger to verify that the view transform definitely changes as the device moves and hence included the inverse of it in the process above.
My code for this experiment (factored into a Unity script) looked like this (pasted in its entirety);
using UnityEngine.XR.WSA;
using System;
using System.Linq;
using UnityEngine;
#if ENABLE_WINMD_SUPPORT
using Windows.Media.Capture;
using Windows.Media.Capture.Frames;
using Windows.Foundation;
using System.Threading.Tasks;
using Windows.Perception.Spatial;
using System.Runtime.InteropServices;
using uVector3 = UnityEngine.Vector3;
using wVector3 = System.Numerics.Vector3;
using wMatrix4x4 = System.Numerics.Matrix4x4;
#endif // ENABLE_WINMD_SUPPORT
public class Placeholder : MonoBehaviour
{
// Unity line renderer to draw a box for me - note that I'm expecting this to have
// Loop set to true so that it closes the box off.
public LineRenderer lineRenderer;
void Start()
{
#if ENABLE_WINMD_SUPPORT
this.OnLoaded();
#endif // ENABLE_WINMD_SUPPORT
}
#if ENABLE_WINMD_SUPPORT
async void OnLoaded()
{
var rgbMedia = await this.GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind.Color, 1280, 720, 30);
// These should be the corner points for the RGB image...
var cornerPoints = new Point[]
{
new Point(0,0),
new Point(1280, 0),
new Point(1280, 720),
new Point(0, 720)
};
var unityWorldCoordinateSystem =
Marshal.GetObjectForIUnknown(WorldManager.GetNativeISpatialCoordinateSystemPtr()) as SpatialCoordinateSystem;
var rgbFrameReader = await rgbMedia.Item1.CreateFrameReaderAsync(rgbMedia.Item2);
rgbFrameReader.AcquisitionMode = MediaFrameReaderAcquisitionMode.Realtime;
TypedEventHandler<MediaFrameReader, MediaFrameArrivedEventArgs> handler =
(s, e) =>
{
using (var frame = s.TryAcquireLatestFrame())
{
if (frame != null)
{
SpatialCoordinateSystem coordinateSystem = null;
wMatrix4x4 projectionTransform = wMatrix4x4.Identity;
wMatrix4x4 viewTransform = wMatrix4x4.Identity;
wMatrix4x4 invertedViewTransform = wMatrix4x4.Identity;
object value;
if (frame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraCoordinateSystem, out value))
{
// I'm not sure that this coordinate system changes per-frame so I could maybe do this once?
coordinateSystem = value as SpatialCoordinateSystem;
}
if (frame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraProjectionTransform, out value))
{
// I don't think that this transform changes per-frame so I could maybe do this once?
projectionTransform = ByteArrayToMatrix(value as byte[]);
}
if (frame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraViewTransform, out value))
{
// I think this transform changes per frame.
viewTransform = ByteArrayToMatrix(value as byte[]);
wMatrix4x4.Invert(viewTransform, out invertedViewTransform);
}
var cameraToWorldTransform = coordinateSystem.TryGetTransformTo(unityWorldCoordinateSystem);
if (cameraToWorldTransform.HasValue)
{
var transformedPoints = cornerPoints
.Select(point => ScalePointMinusOneToOne(point, frame))
.Select(point => UnProjectVector(
new wVector3((float)point.X, (float)point.Y, 1.0f), projectionTransform))
.Select(point => wVector3.Transform(point, invertedViewTransform))
.Select(point => wVector3.Transform(point, cameraToWorldTransform.Value))
.ToArray();
UnityEngine.WSA.Application.InvokeOnAppThread(
() =>
{
this.lineRenderer.positionCount = transformedPoints.Length;
// Unity has Z access +ve away from camera, holographic goes the other way.
this.lineRenderer.SetPositions(
transformedPoints.Select(
pt => new uVector3(pt.X, pt.Y, -1.0f * pt.Z)).ToArray());
},
false);
}
}
}
};
rgbFrameReader.FrameArrived += handler;
await rgbFrameReader.StartAsync();
// Wait forever then dispose...just doing this to keep track of what needs disposing.
await Task.Delay(-1);
rgbFrameReader.FrameArrived -= handler;
Marshal.ReleaseComObject(unityWorldCoordinateSystem);
rgbFrameReader.Dispose();
rgbMedia.Item1.Dispose();
}
/// <summary>
/// This code taken fairly literally from this doc
/// https://docs.microsoft.com/en-us/windows/mixed-reality/locatable-camera#pixel-to-application-specified-coordinate-system
/// and hopefully without me breaking it as it's not too complex 🙂
/// </summary>
static Point ScalePointMinusOneToOne(Point point, MediaFrameReference frameRef)
{
var scaledPoint = new Point(
(2.0f * (float)point.X / frameRef.Format.VideoFormat.Width) - 1.0f,
(2.0f * (1.0f - (float)point.Y / frameRef.Format.VideoFormat.Height)) - 1.0f);
return (scaledPoint);
}
/// <summary>
/// This code taken fairly literally from this doc
/// https://docs.microsoft.com/en-us/windows/mixed-reality/locatable-camera#pixel-to-application-specified-coordinate-system
/// but if it's got messed up in the translation then that's definitely my fault 🙂
/// </summary>
static wVector3 UnProjectVector(wVector3 from, wMatrix4x4 cameraProjection)
{
var to = new wVector3(0, 0, 0);
var axsX = new wVector3(cameraProjection.M11, cameraProjection.M12, cameraProjection.M13);
var axsY = new wVector3(cameraProjection.M21, cameraProjection.M22, cameraProjection.M23);
var axsZ = new wVector3(cameraProjection.M31, cameraProjection.M32, cameraProjection.M33);
to.Z = from.Z / axsZ.Z;
to.Y = (from.Y - (to.Z * axsY.Z)) / axsY.Y;
to.X = (from.X - (to.Z * axsX.Z)) / axsX.X;
return to;
}
// Used an explicit tuple here as I'm in C# 6.0
async Task<Tuple<MediaCapture, MediaFrameSource>> GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind sourceKind,
int width,
int height,
int frameRate)
{
MediaCapture mediaCapture = null;
MediaFrameSource frameSource = null;
var allSources = await MediaFrameSourceGroup.FindAllAsync();
var sourceInfo =
allSources.SelectMany(group => group.SourceInfos)
.FirstOrDefault(
si =>
(si.MediaStreamType == MediaStreamType.VideoRecord) &&
(si.SourceKind == sourceKind) &&
(si.VideoProfileMediaDescription.Any(
desc =>
desc.Width == width &&
desc.Height == height &&
desc.FrameRate == frameRate)));
if (sourceInfo != null)
{
var sourceGroup = sourceInfo.SourceGroup;
mediaCapture = new MediaCapture();
await mediaCapture.InitializeAsync(
new MediaCaptureInitializationSettings()
{
// I want software bitmaps
MemoryPreference = MediaCaptureMemoryPreference.Cpu,
SourceGroup = sourceGroup,
StreamingCaptureMode = StreamingCaptureMode.Video
}
);
frameSource = mediaCapture.FrameSources[sourceInfo.Id];
var selectedFormat = frameSource.SupportedFormats.First(
format => format.VideoFormat.Width == width && format.VideoFormat.Height == height &&
format.FrameRate.Numerator / format.FrameRate.Denominator == frameRate);
await frameSource.SetFormatAsync(selectedFormat);
}
return (Tuple.Create(mediaCapture, frameSource));
}
static wMatrix4x4 ByteArrayToMatrix(byte[] bits)
{
var matrix = wMatrix4x4.Identity;
var handle = GCHandle.Alloc(bits, GCHandleType.Pinned);
matrix = Marshal.PtrToStructure<wMatrix4x4>(handle.AddrOfPinnedObject());
handle.Free();
return (matrix);
}
static readonly Guid MFSampleExtension_Spatial_CameraCoordinateSystem = new Guid("9D13C82F-2199-4E67-91CD-D1A4181F2534");
static readonly Guid MFSampleExtension_Spatial_CameraProjectionTransform = new Guid("47F9FCB5-2A02-4F26-A477-792FDF95886A");
static readonly Guid MFSampleExtension_Spatial_CameraViewTransform = new Guid("4E251FA4-830F-4770-859A-4B8D99AA809B");
#endif // ENABLE_WINMD_SUPPORT
}
and this seemed to work out ok in the sense that I could run this code on my HoloLens and see a painted red line demarcating what ‘felt’ like it might be the right positions of the camera’s view and that box appeared to do the right thing as I moved around and rotated the device etc. but I wouldn’t have placed (much) money on it being correct just yet
Experiment 7 – Mapping Between RGB Co-ordinates and Depth Co-Ordinates
The last experiment that I wanted to try was to see if I could figure out how to map co-ordinates from the RGB image to the depth image.
On the one hand, this seems like it might be ‘obvious’ in the sense that if I have a pixel at some X,Y in an RGB image [0,0,1280,720] and if I have some depth image which is 448×450 then I can just come up with a point which is [ X / 1280 *448, Y / 720 * 450] and use that as the position in the depth image.
However, I don’t know whether the depth image is meant to line up with the RGB image in that way or whether I should use other techniques in trying to map depth image coordinates to/from RGB image coordinates.
While I was experimenting with this, an additional sample was published around working with ‘research mode’ and so I was able to refer to it;
https://github.com/Microsoft/HoloLensForCV
and, firstly, I found that I had to make a minor modification to it in the FrameRenderer.cpp code that it uses because around line 348 it hard-codes the depth range to 0.2m to 1.0m whereas I want ranges beyond 1.0m.
NB: I think that has now been addressed – see this issue.
With that modification made, I saw the output from the depth camera as below;
which seems to suggest that the depth values that I want aren’t present across the whole frame (448 x 450) of the depth image but look to be, instead, present in a circular area which you can see highlighted above.
That also seems to be the case for the “Long Throw ToF Reflectivity” stream and I can speculate that maybe those sensors focus (for power/performance reasons?) around the centre of the user’s gaze but that’s just speculation, I don’t see that written down anywhere at the time of writing.
Furthermore, that circular area does not seem to line up with the centre of the depth frame. For instance, in the image below;
my gaze is on the corner of the book-case marked with a green X which looks to be fairly centrally located in the RGB camera captured image here but the centre of the depth frame seems biased towards the top of the frame and so I can’t simply assume that I can scale coordinates from the RGB frame to the depth frame and come away with reasonable depth values.
This made my original idea seem a lot less practical than it might have seemed when I first started writing this post because I’d assumed that every RGB camera pixel would have a natural corresponding depth camera pixel and I’m not sure whether that’s going to be the case.
So, perhaps the depth camera is better for working out depths around where the user’s gaze is positioned (which makes sense) and my facial example is then only realistically going to ‘work’ if the user is looking directly at a face.
Additionally, even if I assume that I want to measure the depth value at the centre of the RGB image [640,360] then I can’t assume that this maps to the co-ordinate [224, 225] in the depth image because the depth image seems to incorporate a vertical offset.
Or…maybe I’m just missing quite a lot about how these streams can be tied together?
I wanted to see what did happen if I brought the pieces together that I had so far and so I tried to put together a Unity script which moves a GameObject (e.g. small sphere) to the centre point of any face that it detects in the RGB stream.
That script is below (it needs factoring out into classes as it’s mostly one large function at the moment);
//#define HUNT_DEPTH_PIXEL_GRID
#define USE_CENTRE_DEPTH_IMAGE
using UnityEngine.XR.WSA;
using System;
using System.Linq;
using UnityEngine;
using System.Threading;
#if ENABLE_WINMD_SUPPORT
using Windows.Media.Capture;
using Windows.Media.Capture.Frames;
using Windows.Foundation;
using System.Threading.Tasks;
using Windows.Perception.Spatial;
using System.Runtime.InteropServices;
using Windows.Media.FaceAnalysis;
using Windows.Graphics.Imaging;
using uVector3 = UnityEngine.Vector3;
using wVector3 = System.Numerics.Vector3;
using wVector4 = System.Numerics.Vector4;
using wMatrix4x4 = System.Numerics.Matrix4x4;
[ComImport]
[Guid("5B0D3235-4DBA-4D44-865E-8F1D0E4FD04D")]
[InterfaceType(ComInterfaceType.InterfaceIsIUnknown)]
unsafe interface IMemoryBufferByteAccess
{
void GetBuffer(out byte* buffer, out uint capacity);
}
#endif // ENABLE_WINMD_SUPPORT
public class Placeholder : MonoBehaviour
{
// A Unity text mesh that I can print some diagnostics to.
public TextMesh textMesh;
// A Unity game object (small sphere e.g.) that I can use to mark the position of one face.
public GameObject faceMarker;
void Start()
{
#if ENABLE_WINMD_SUPPORT
// Not awaiting this...let it go.
this.ProcessingLoopAsync();
#endif // ENABLE_WINMD_SUPPORT
}
#if ENABLE_WINMD_SUPPORT
/// <summary>
/// This is just one big lump of code right now which should be factored out into some kind of
/// 'frame reader' class which can then be subclassed for depth frame and video frame but
/// it was handy to have it like this while I experimented with it - the intention was
/// to tidy it up if I could get it doing more or less what I wanted 🙂
/// </summary>
async Task ProcessingLoopAsync()
{
var depthMediaCapture = await this.GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind.Depth, 448, 450, 15);
var depthFrameReader = await depthMediaCapture.Item1.CreateFrameReaderAsync(depthMediaCapture.Item2);
depthFrameReader.AcquisitionMode = MediaFrameReaderAcquisitionMode.Realtime;
MediaFrameReference lastDepthFrame = null;
long depthFrameCount = 0;
float centrePointDepthInMetres = 0.0f;
// Expecting this to run at 1fps although the API (seems to) reports that it runs at 15fps
TypedEventHandler<MediaFrameReader, MediaFrameArrivedEventArgs> depthFrameHandler =
(sender, args) =>
{
using (var depthFrame = sender.TryAcquireLatestFrame())
{
if ((depthFrame != null) && (depthFrame != lastDepthFrame))
{
lastDepthFrame = depthFrame;
Interlocked.Increment(ref depthFrameCount);
// Always try to grab the depth value although, clearly, this is subject
// to a bunch of race conditions etc. as other thread access it.
centrePointDepthInMetres =
GetDepthValueAtCoordinate(depthFrame,
(int)(depthFrame.Format.VideoFormat.Width * MAGIC_DEPTH_FRAME_WIDTH_RATIO_CENTRE),
(int)(depthFrame.Format.VideoFormat.Height * MAGIC_DEPTH_FRAME_HEIGHT_RATIO_CENTRE)) ?? 0.0f;
}
}
};
long rgbProcessedCount = 0;
long facesPresentCount = 0;
long rgbDroppedCount = 0;
MediaFrameReference lastRgbFrame = null;
var faceBitmapFormats = FaceTracker.GetSupportedBitmapPixelFormats().Select(
format => format.ToString().ToLower()).ToArray();
var faceTracker = await FaceTracker.CreateAsync();
var rgbMediaCapture = await this.GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind.Color, 1280, 720, 30, faceBitmapFormats);
var rgbFrameReader = await rgbMediaCapture.Item1.CreateFrameReaderAsync(rgbMediaCapture.Item2);
rgbFrameReader.AcquisitionMode = MediaFrameReaderAcquisitionMode.Realtime;
int busyProcessingRgbFrame = 0;
var unityWorldCoordinateSystem =
Marshal.GetObjectForIUnknown(WorldManager.GetNativeISpatialCoordinateSystemPtr()) as SpatialCoordinateSystem;
// Expecting this to run at 30fps.
TypedEventHandler<MediaFrameReader, MediaFrameArrivedEventArgs> rgbFrameHandler =
(sender, args) =>
{
// Only proceed if we're not already 'busy' - i.e. we'
if (Interlocked.CompareExchange(ref busyProcessingRgbFrame, 1, 0) == 0)
{
Task.Run(
async () =>
{
using (var rgbFrame = rgbFrameReader.TryAcquireLatestFrame())
{
if ((rgbFrame != null) && (rgbFrame != lastRgbFrame))
{
++rgbProcessedCount;
lastRgbFrame = rgbFrame;
var facePosition = uVector3.zero;
using (var videoFrame = rgbFrame.VideoMediaFrame.GetVideoFrame())
{
var faces = await faceTracker.ProcessNextFrameAsync(videoFrame);
var firstFace = faces.FirstOrDefault();
if (firstFace != null)
{
++facesPresentCount;
// Take the first face and the centre point of that face to try
// and simplify things for my limited brain.
var faceCentrePointInImageCoords =
new Point(
firstFace.FaceBox.X + (firstFace.FaceBox.Width / 2.0f),
firstFace.FaceBox.Y + (firstFace.FaceBox.Height / 2.0f));
wMatrix4x4 projectionTransform = wMatrix4x4.Identity;
wMatrix4x4 viewTransform = wMatrix4x4.Identity;
wMatrix4x4 invertedViewTransform = wMatrix4x4.Identity;
var rgbCoordinateSystem = GetRgbFrameProjectionAndCoordinateSystemDetails(
rgbFrame, out projectionTransform, out invertedViewTransform);
// Scale the RGB point (1280x720)
var faceCentrePointUnitScaleRGB = ScalePointMinusOneToOne(faceCentrePointInImageCoords, rgbFrame);
// Unproject the RGB point back at unit depth as per the locatable camera
// document.
var unprojectedFaceCentrePointRGB = UnProjectVector(
new wVector3(
(float)faceCentrePointUnitScaleRGB.X,
(float)faceCentrePointUnitScaleRGB.Y,
1.0f),
projectionTransform);
// Transfrom this back by the inverted view matrix in order to put this into
// the RGB camera coordinate system
var faceCentrePointCameraCoordsRGB =
wVector3.Transform(unprojectedFaceCentrePointRGB, invertedViewTransform);
// Get the transform from the camera coordinate system to the Unity world
// coordinate system, could probably cache this?
var cameraRGBToWorldTransform =
rgbCoordinateSystem.TryGetTransformTo(unityWorldCoordinateSystem);
if (cameraRGBToWorldTransform.HasValue)
{
// Transform to world coordinates
var faceCentrePointWorldCoords = wVector4.Transform(
new wVector4(
faceCentrePointCameraCoordsRGB.X,
faceCentrePointCameraCoordsRGB.Y,
faceCentrePointCameraCoordsRGB.Z, 1),
cameraRGBToWorldTransform.Value);
// Where's the camera in world coordinates?
var cameraOriginWorldCoords = wVector4.Transform(
new wVector4(0, 0, 0, 1),
cameraRGBToWorldTransform.Value);
// Multiply Z by -1 for Unity
var cameraPoint = new uVector3(
cameraOriginWorldCoords.X,
cameraOriginWorldCoords.Y,
-1.0f * cameraOriginWorldCoords.Z);
// Multiply Z by -1 for Unity
var facePoint = new uVector3(
faceCentrePointWorldCoords.X,
faceCentrePointWorldCoords.Y,
-1.0f * faceCentrePointWorldCoords.Z);
facePosition =
cameraPoint +
(facePoint - cameraPoint).normalized * centrePointDepthInMetres;
}
}
}
if (facePosition != uVector3.zero)
{
UnityEngine.WSA.Application.InvokeOnAppThread(
() =>
{
this.faceMarker.transform.position = facePosition;
},
false
);
}
}
}
Interlocked.Exchange(ref busyProcessingRgbFrame, 0);
}
);
}
else
{
Interlocked.Increment(ref rgbDroppedCount);
}
// NB: this is a bit naughty as I am accessing these counters across a few threads so
// accuracy might suffer here.
UnityEngine.WSA.Application.InvokeOnAppThread(
() =>
{
this.textMesh.text =
$"{depthFrameCount} depth,{rgbProcessedCount} rgb done, {rgbDroppedCount} rgb drop," +
$"{facesPresentCount} faces, ({centrePointDepthInMetres:N2})";
},
false);
};
depthFrameReader.FrameArrived += depthFrameHandler;
rgbFrameReader.FrameArrived += rgbFrameHandler;
await depthFrameReader.StartAsync();
await rgbFrameReader.StartAsync();
// Wait forever then dispose...just doing this to keep track of what needs disposing.
await Task.Delay(-1);
depthFrameReader.FrameArrived -= depthFrameHandler;
rgbFrameReader.FrameArrived -= rgbFrameHandler;
rgbFrameReader.Dispose();
depthFrameReader.Dispose();
rgbMediaCapture.Item1.Dispose();
depthMediaCapture.Item1.Dispose();
Marshal.ReleaseComObject(unityWorldCoordinateSystem);
}
static SpatialCoordinateSystem GetRgbFrameProjectionAndCoordinateSystemDetails(
MediaFrameReference rgbFrame,
out wMatrix4x4 projectionTransform,
out wMatrix4x4 invertedViewTransform)
{
SpatialCoordinateSystem rgbCoordinateSystem = null;
wMatrix4x4 viewTransform = wMatrix4x4.Identity;
projectionTransform = wMatrix4x4.Identity;
invertedViewTransform = wMatrix4x4.Identity;
object value;
if (rgbFrame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraCoordinateSystem, out value))
{
// I'm not sure that this coordinate system changes per-frame so I could maybe do this once?
rgbCoordinateSystem = value as SpatialCoordinateSystem;
}
if (rgbFrame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraProjectionTransform, out value))
{
// I don't think that this transform changes per-frame so I could maybe do this once?
projectionTransform = ByteArrayToMatrix(value as byte[]);
}
if (rgbFrame.Properties.TryGetValue(MFSampleExtension_Spatial_CameraViewTransform, out value))
{
// I think this transform changes per frame.
viewTransform = ByteArrayToMatrix(value as byte[]);
wMatrix4x4.Invert(viewTransform, out invertedViewTransform);
}
return (rgbCoordinateSystem);
}
/// <summary>
/// Not using this right now as I don't *know* how to scale an RGB point to a depth point
/// given that the depth frame seems to have a central 'hot spot' that's circular.
/// </summary>
static Point ScaleRgbPointToDepthPoint(Point rgbPoint, MediaFrameReference rgbFrame,
MediaFrameReference depthFrame)
{
return (new Point(
rgbPoint.X / rgbFrame.Format.VideoFormat.Width * depthFrame.Format.VideoFormat.Width,
rgbPoint.Y / rgbFrame.Format.VideoFormat.Height * depthFrame.Format.VideoFormat.Height));
}
/// <summary>
/// This code taken fairly literally from this doc
/// https://docs.microsoft.com/en-us/windows/mixed-reality/locatable-camera#pixel-to-application-specified-coordinate-system
/// and hopefully without me breaking it too badly.
/// </summary>
static Point ScalePointMinusOneToOne(Point point, MediaFrameReference frameRef)
{
var scaledPoint = new Point(
(2.0f * (float)point.X / frameRef.Format.VideoFormat.Width) - 1.0f,
(2.0f * (1.0f - (float)point.Y / frameRef.Format.VideoFormat.Height)) - 1.0f);
return (scaledPoint);
}
/// <summary>
/// This code taken fairly literally from this doc
/// https://docs.microsoft.com/en-us/windows/mixed-reality/locatable-camera#pixel-to-application-specified-coordinate-system
/// but if it's got messed up in the translation then that's definitely my fault 🙂
/// </summary>
static wVector3 UnProjectVector(wVector3 from, wMatrix4x4 cameraProjection)
{
var to = new wVector3(0, 0, 0);
var axsX = new wVector3(cameraProjection.M11, cameraProjection.M12, cameraProjection.M13);
var axsY = new wVector3(cameraProjection.M21, cameraProjection.M22, cameraProjection.M23);
var axsZ = new wVector3(cameraProjection.M31, cameraProjection.M32, cameraProjection.M33);
to.Z = from.Z / axsZ.Z;
to.Y = (from.Y - (to.Z * axsY.Z)) / axsY.Y;
to.X = (from.X - (to.Z * axsX.Z)) / axsX.X;
return to;
}
unsafe static float? GetDepthValueAtCoordinate(MediaFrameReference frame, int x, int y)
{
float? depthValue = null;
var bitmap = frame.VideoMediaFrame.SoftwareBitmap;
using (var buffer = bitmap.LockBuffer(BitmapBufferAccessMode.Read))
using (var reference = buffer.CreateReference())
{
var description = buffer.GetPlaneDescription(0);
byte* pBits;
uint size;
var byteAccess = reference as IMemoryBufferByteAccess;
byteAccess.GetBuffer(out pBits, out size);
// Try the pixel value itself and see if we get anything there.
depthValue = GetDepthValueFromBufferAtXY(
pBits, x, y, description, (float)frame.VideoMediaFrame.DepthMediaFrame.DepthFormat.DepthScaleInMeters);
#if HUNT_DEPTH_PIXEL_GRID
if (depthValue == null)
{
// If we don't have a value, look for one in the surrounding space (the sub-function copes
// with us using bad values of x,y).
var minDistance = double.MaxValue;
for (int i = 0 - DEPTH_SEARCH_GRID_SIZE; i < DEPTH_SEARCH_GRID_SIZE; i++)
{
for (int j = 0 - DEPTH_SEARCH_GRID_SIZE; j < DEPTH_SEARCH_GRID_SIZE; j++)
{
var newX = x + i;
var newY = y + j;
var testValue = GetDepthValueFromBufferAtXY(
pBits,
newX,
newY,
description,
(float)frame.VideoMediaFrame.DepthMediaFrame.DepthFormat.DepthScaleInMeters);
if (testValue != null)
{
var distance =
Math.Sqrt(Math.Pow(newX - x, 2.0) + Math.Pow(newY - y, 2.0));
if (distance < minDistance)
{
depthValue = testValue;
minDistance = distance;
}
}
}
}
}
#endif // HUNT_DEPTH_PIXEL_GRID
}
return (depthValue);
}
unsafe static float? GetDepthValueFromBufferAtXY(byte* pBits, int x, int y, BitmapPlaneDescription desc,
float scaleInMeters)
{
float? depthValue = null;
var bytesPerPixel = desc.Stride / desc.Width;
Debug.Assert(bytesPerPixel == Marshal.SizeOf<UInt16>());
int offset = (desc.StartIndex + desc.Stride * y) + (x * bytesPerPixel);
if ((offset > 0) && (offset < ((long)pBits + (desc.Height * desc.Stride))))
{
depthValue = *(UInt16*)(pBits + offset) * scaleInMeters;
if (!IsValidDepthDistance((float)depthValue))
{
depthValue = null;
}
}
return (depthValue);
}
static bool IsValidDepthDistance(float depthDistance)
{
// If that depth value is > 4.0m then we discard it because it seems like
// 4.**m (4.09?) comes back from the sensor when it hasn't really got a value
return ((depthDistance > 0.5f) && (depthDistance <= 4.0f));
}
// Used an explicit tuple here as I'm in C# 6.0
async Task<Tuple<MediaCapture, MediaFrameSource>> GetMediaCaptureForDescriptionAsync(
MediaFrameSourceKind sourceKind,
int width,
int height,
int frameRate,
string[] bitmapFormats = null)
{
MediaCapture mediaCapture = null;
MediaFrameSource frameSource = null;
var allSources = await MediaFrameSourceGroup.FindAllAsync();
// Ignore frame rate here on the description as both depth streams seem to tell me they are
// 30fps whereas I don't think they are (from the docs) so I leave that to query later on.
// NB: LastOrDefault here is a NASTY, NASTY hack - just my way of getting hold of the
// *LAST* depth stream rather than the *FIRST* because I'm assuming that the *LAST*
// one is the longer distance stream rather than the short distance stream.
// I should fix this and find a better way of choosing the right depth stream rather
// than relying on some ordering that's not likely to always work!
var sourceInfo =
allSources.SelectMany(group => group.SourceInfos)
.LastOrDefault(
si =>
(si.MediaStreamType == MediaStreamType.VideoRecord) &&
(si.SourceKind == sourceKind) &&
(si.VideoProfileMediaDescription.Any(
desc =>
desc.Width == width &&
desc.Height == height &&
desc.FrameRate == frameRate)));
if (sourceInfo != null)
{
var sourceGroup = sourceInfo.SourceGroup;
mediaCapture = new MediaCapture();
await mediaCapture.InitializeAsync(
new MediaCaptureInitializationSettings()
{
// I want software bitmaps
MemoryPreference = MediaCaptureMemoryPreference.Cpu,
SourceGroup = sourceGroup,
StreamingCaptureMode = StreamingCaptureMode.Video
}
);
frameSource = mediaCapture.FrameSources[sourceInfo.Id];
var selectedFormat = frameSource.SupportedFormats.First(
format =>
format.VideoFormat.Width == width && format.VideoFormat.Height == height &&
format.FrameRate.Numerator / format.FrameRate.Denominator == frameRate &&
((bitmapFormats == null) || (bitmapFormats.Contains(format.Subtype.ToLower()))));
await frameSource.SetFormatAsync(selectedFormat);
}
return (Tuple.Create(mediaCapture, frameSource));
}
static wMatrix4x4 ByteArrayToMatrix(byte[] bits)
{
var matrix = wMatrix4x4.Identity;
var handle = GCHandle.Alloc(bits, GCHandleType.Pinned);
matrix = Marshal.PtrToStructure<wMatrix4x4>(handle.AddrOfPinnedObject());
handle.Free();
return (matrix);
}
#if HUNT_DEPTH_PIXEL_GRID
static readonly int DEPTH_SEARCH_GRID_SIZE = 32;
#endif // HUNT_DEPTH_PIXEL_GRID
static readonly float MAGIC_DEPTH_FRAME_HEIGHT_RATIO_CENTRE = 0.25f;
static readonly float MAGIC_DEPTH_FRAME_WIDTH_RATIO_CENTRE = 0.5f;
static readonly Guid MFSampleExtension_Spatial_CameraCoordinateSystem = new Guid("9D13C82F-2199-4E67-91CD-D1A4181F2534");
static readonly Guid MFSampleExtension_Spatial_CameraProjectionTransform = new Guid("47F9FCB5-2A02-4F26-A477-792FDF95886A");
static readonly Guid MFSampleExtension_Spatial_CameraViewTransform = new Guid("4E251FA4-830F-4770-859A-4B8D99AA809B");
#endif // ENABLE_WINMD_SUPPORT
}
and it produces a sort of bouncing ball which (ideally) hovers around faces as shown in the screen capture below which makes it look better than it actually is at finding faces
with some on-screen diagnostics trying to show how many;
- depth frames we have seen
- RGB frames we have seen
- RGB frames we have ignored because we were still processing the previous frame
- frames we have seen which contained faces
along with the current depth value obtained from the ‘centre point’ of the camera which you’ll notice in the code is hard-coded to be a point 25% down the frame and 50% across – that’s just a ‘best guess’ right now rather than anything ‘scientific’.
Wrapping Up the Experiments for Now
I clearly need to spend some more time experimenting here as I haven’t quite got to the result that I wanted to but I learned quite a lot along the way even if my results might be a little flawed.
Through this post, I’ve been questioning my initial assumption that using the depth frames for estimating the Z-coordinate of a face was a good route to take.
Maybe that’s not right? Given that I get a long-range depth frame at 1fps and given that the depth data seems to be concentrated in one area of that frame, perhaps it doesn’t make sense to try to use the depth frame in this way to identify the Z-coordinate of a face (or other object in space). Maybe it’s better to go via the regular route of using the spatial mesh which the device builds so well after all?
I need to try a few more things out
Code?
I haven’t published separate pieces of code for all of the experiments above but the Unity project that I have which brought some of them together in the last experiment is in this repo;
https://github.com/mtaulty/FacialDepthExperiments
Note that if you take the code and build it from Unity then you’ll need to mark the C# project assembly as allowing unsafe code before you’ll be able to get Visual Studio to build it – I can’t find a setting in Unity that seems to allow unsafe code in the separate C# project assembly rather than the main executable itself.
Note also that you will need to manually edit the .appxmanifest file to add the restricted persmission for perceptionSensorsExperimental as I wrote up in a previous post because there is no way (as far as I know) to set this in either the Unity or Visual Studio editors.
Lastly, apply a pinch of salt – I’m just experimenting here
