Get ByteBuffer from Image for TensorFlow Lite Model - java

I am creating an android app to run on Google Glass Enterprise Edition 2 that does Real-time Face Recognition. I am using Camera X as my Camera API and TensorFlow Lite (TFLite) as my classification model. However, the TFLite model input requires ByteBuffer which I am unable to convert into from the image retrieved from CameraX.
How do I get my Image from CameraX into ByteBuffer class for my TFLite Model?
Camera X Image Analysis: Reference
val imageAnalysis = ImageAnalysis.Builder()
.setTargetResolution(Size(640, 360))
.setBackpressureStrategy(ImageAnalysis.STRATEGY_KEEP_ONLY_LATEST)
.build()
imageAnalysis.setAnalyzer(AsyncTask.THREAD_POOL_EXECUTOR, ImageAnalysis.Analyzer { imageProxy ->
val rotationDegrees = imageProxy.imageInfo.rotationDegrees
val mediaImage = imageProxy.image
if (mediaImage != null) {
val image = InputImage.fromMediaImage(mediaImage, rotationDegrees)
/* Classify the Image using TensorFlow Lite Model */
}
})
TensorFlow Model Sample Code
val model = FaceRecognitionModel.newInstance(context)
// Creates inputs for reference.
val inputFeature0 = TensorBuffer.createFixedSize(intArrayOf(1, 224, 224, 3), DataType.FLOAT32)
inputFeature0.loadBuffer(byteBuffer)
// Runs model inference and gets result.
val outputs = model.process(inputFeature0)
val outputFeature0 = outputs.outputFeature0AsTensorBuffer
// Releases model resources if no longer used.
model.close()

Try using the TensorImage class from the TensorFlow Lite Support Library.
Roughly, you can follow these steps.
Convert the Image object into Bitmap. There should be other Stackoverflow questions on how to do this (e.g., this answer)
Create a TensorImage object from the Bitmap object using TensorImage.fromBitmap() factory.
Call getBuffer() method on the TensorImage object to get the underlying ByteBuffer.
You might also want to do some image pre-processing, in case the image from CameraX doesn't exactly match the format expected by the model. For this, you can explore the ImageProcessor utility.

I have made some findings and I applied these to help shorten my problem.
The image I have gotten from CameraX is in YUV. I have trained my model in RGB in 224X224. To suit my issue, I first convert the image to RGB Bitmap, then crop it into 224X224. Afterwards convert Bitmap to ByteBuffer.
As for my TFLite Model, the TFLite model managed to accept the converted RGB ByteBuffer image and process it. Returning back
TensorBuffer.

Related

Android Camera2 Preview Image Decoding fails on car head unit

I am pretty new to Camera2 API and currenty trying to implement a camera preview into my own car application on an Android head unit (Model YT9213AJ). The preview should show the image of the reverse camera.
I've tested the following code on a Samsung Tablet (SM-P610) and it shows the camera preview images as expected, from both, rear and front camera.
private void bindPreview(#NonNull ProcessCameraProvider cameraProvider) {
CameraSelector cameraSelector = new CameraSelector.Builder()
.requireLensFacing(CameraSelector.LENS_FACING_FRONT)
.build();
CameraManager manager = (CameraManager) context.getSystemService(CAMERA_SERVICE);
Size previewSize = getPreviewSize(manager.getCameraCharacteristics("1"));
Preview preview = new Preview.Builder()
.setTargetResolution(previewSize)
.setDefaultResolution(previewSize)
.setMaxResolution(previewSize)
.setTargetRotation(Surface.ROTATION_270)
.build();
preview.setSurfaceProvider(previewView.getSurfaceProvider());
Camera camera = cameraProvider.bindToLifecycle((LifecycleOwner)this, cameraSelector, preview);
}
With this function to get the preview size:
Size getPreviewSize(CameraCharacteristics characteristics) {
StreamConfigurationMap map = characteristics.get(CameraCharacteristics.SCALER_STREAM_CONFIGURATION_MAP);
Size[] previewSizes = map.getOutputSizes(SurfaceTexture.class);
return previewSizes[0]; // for the camera just one resolution is given, so this should be sufficient
}
This is what the image looks like, if I run it on the car head unit:
Example image preview from car head unit
(Sorry, can't embed images into post yet)
I've also run the App "Camera2 API probe", please find the results here on AirBeat. The camera with ID 5 seems to be an placeholder, I assume that the other two cameras (ID 0 and ID 1) are the representations of the two hardware inputs into the head unit.
Do you have any clue how I can correctly decode the image for this camera model? Thanks for your time.
You could try seeing if not rotating the output helps - the size listed by the camera is already portrait aspect ratio.
In general, though, it looks like whoever made this head unit didn't take a lot of care in making sure the camera output works with the display side, I'm guessing. So it may not be possible to get it to work with the CameraX PreviewView as-is. The PreviewView tries to select the best kind of output View for the API level that the device is, but it may be picking a path the manufacturer didn't actually test with. There's no built-in way to have it select the other option (SurfaceView or TextureView are the choices), so you'd have to modify PreviewView code or build your own Preview surface provider.
There's also no guarantee that'll help, depending on exactly how badly this device implements things.

Custom video source for WebRTC on Android

Overview
I would like to use a custom video source to live stream video via WebRTC Android implementation. If I understand correctly, existing implementation only supports front and back facing cameras on Android phones. The following classes are relevant in this scenario:
Camera1Enumerator.java
VideoCapturer.java
PeerConnectionFactory
VideoSource.java
VideoTrack.java
Currently for using front facing camera on Android phone I'm doing the following steps:
CameraEnumerator enumerator = new Camera1Enumerator(false);
VideoCapturer videoCapturer = enumerator.createCapturer(deviceName, null);
VideoSource videoSource = peerConnectionFactory.createVideoSource(false);
videoCapturer.initialize(surfaceTextureHelper, this.getApplicationContext(), videoSource.getCapturerObserver());
VideoTrack localVideoTrack = peerConnectionFactory.createVideoTrack(VideoTrackID, videoSource);
My scenario
I've a callback handler that receives video buffer in byte array from custom video source:
public void onReceive(byte[] videoBuffer, int size) {}
How would I be able to send this byte array buffer? I'm not sure about the solution, but I think I would have to implement custom VideoCapturer?
Existing questions
This question might be relevant, though I'm not using libjingle library, only native WebRTC Android package.
Similar questions/articles:
for iOS platform but unfortunately I couldn't help with the answers.
for native C++ platform
article about native implementation
There are two possible solutions to this problem:
Implement custom VideoCapturer and create VideoFrame using byte[] stream data in onReceive handler. There actually exists a very good example of FileVideoCapturer, which implements VideoCapturer.
Simply construct VideoFrame from NV21Buffer, which is created from our byte array stream data. Then we only need to use our previously created VideoSource to capture this frame. Example:
public void onReceive(byte[] videoBuffer, int size, int width, int height) {
long timestampNS = TimeUnit.MILLISECONDS.toNanos(SystemClock.elapsedRealtime());
NV21Buffer buffer = new NV21Buffer(videoBuffer, width, height, null);
VideoFrame videoFrame = new VideoFrame(buffer, 0, timestampNS);
videoSource.getCapturerObserver().onFrameCaptured(videoFrame);
videoFrame.release();
}

How to implement the function of getYourInputImage()?

I'm working on eye detection project and I have trained my own model using Tensorflow Library, I have now the Tensorflow lite model.
I followed the documentation proposed by Tensorflow for how to integrate custom model MLKit firebase on the Android platform.
This is CustomModelActivity.java that I followed
https://github.com/firebase/snippets-android/blob/a7be1d9b6208d112d1bacbfff937c32c219124b5/mlkit/app/src/main/java/com/google/firebase/example/mlkit/CustomModelActivity.java#L127-L128
I tried to implement this function by setting the image on imageview and I have tried to convert it to bitmap
But I have encountered some issue ( false predictions ) on an image from the dataset.
I have posted many questions about this problem but I didn't get any answer.
I would like now to load the image from my gallery and get it, My question is:
How can I implement the last function getYouInputImage() ?
private Bitmap getYourInputImage() {
// This method is just for show
return Bitmap.createBitmap(0, 0, Bitmap.Config.ALPHA_8);
}

Android Java - programmatically capture background before app window opens [duplicate]

This question already has answers here:
Closed 10 years ago.
Possible Duplicate:
How to programatically take a screenshot on Android?
How to capture the android device screen content and make an image file using the snapshot data? Which API should I use or where could I find related resources?
BTW:
not camera snapshot, but device screen
Use the following code:
Bitmap bitmap;
View v1 = MyView.getRootView();
v1.setDrawingCacheEnabled(true);
bitmap = Bitmap.createBitmap(v1.getDrawingCache());
v1.setDrawingCacheEnabled(false);
Here MyView is the View through which we need include in the screen. You can also get DrawingCache from of any View this way (without getRootView()).
There is also another way.. If we having ScrollView as root view then its better to use following code,
LayoutInflater inflater = (LayoutInflater) this.getSystemService(LAYOUT_INFLATER_SERVICE);
FrameLayout root = (FrameLayout) inflater.inflate(R.layout.activity_main, null); // activity_main is UI(xml) file we used in our Activity class. FrameLayout is root view of my UI(xml) file.
root.setDrawingCacheEnabled(true);
Bitmap bitmap = getBitmapFromView(this.getWindow().findViewById(R.id.frameLayout)); // here give id of our root layout (here its my FrameLayout's id)
root.setDrawingCacheEnabled(false);
Here is the getBitmapFromView() method
public static Bitmap getBitmapFromView(View view) {
//Define a bitmap with the same size as the view
Bitmap returnedBitmap = Bitmap.createBitmap(view.getWidth(), view.getHeight(),Bitmap.Config.ARGB_8888);
//Bind a canvas to it
Canvas canvas = new Canvas(returnedBitmap);
//Get the view's background
Drawable bgDrawable =view.getBackground();
if (bgDrawable!=null)
//has background drawable, then draw it on the canvas
bgDrawable.draw(canvas);
else
//does not have background drawable, then draw white background on the canvas
canvas.drawColor(Color.WHITE);
// draw the view on the canvas
view.draw(canvas);
//return the bitmap
return returnedBitmap;
}
It will display entire screen including content hidden in your ScrollView
UPDATED AS ON 20-04-2016
There is another better way to take screenshot.Here I have taken screenshot of WebView.
WebView w = new WebView(this);
w.setWebViewClient(new WebViewClient()
{
public void onPageFinished(final WebView webView, String url) {
new Handler().postDelayed(new Runnable(){
#Override
public void run() {
webView.measure(View.MeasureSpec.makeMeasureSpec(
View.MeasureSpec.UNSPECIFIED, View.MeasureSpec.UNSPECIFIED),
View.MeasureSpec.makeMeasureSpec(0, View.MeasureSpec.UNSPECIFIED));
webView.layout(0, 0, webView.getMeasuredWidth(),
webView.getMeasuredHeight());
webView.setDrawingCacheEnabled(true);
webView.buildDrawingCache();
Bitmap bitmap = Bitmap.createBitmap(webView.getMeasuredWidth(),
webView.getMeasuredHeight(), Bitmap.Config.ARGB_8888);
Canvas canvas = new Canvas(bitmap);
Paint paint = new Paint();
int height = bitmap.getHeight();
canvas.drawBitmap(bitmap, 0, height, paint);
webView.draw(canvas);
if (bitmap != null) {
try {
String filePath = Environment.getExternalStorageDirectory()
.toString();
OutputStream out = null;
File file = new File(filePath, "/webviewScreenShot.png");
out = new FileOutputStream(file);
bitmap.compress(Bitmap.CompressFormat.PNG, 50, out);
out.flush();
out.close();
bitmap.recycle();
} catch (Exception e) {
e.printStackTrace();
}
}
}
}, 1000);
}
});
Hope this helps..!
AFAIK, All of the methods currently to capture a screenshot of android use the /dev/graphics/fb0 framebuffer. This includes ddms. It does require root to read from this stream. ddms uses adbd to request the information, so root is not required as adb has the permissions needed to request the data from /dev/graphics/fb0.
The framebuffer contains 2+ "frames" of RGB565 images. If you are able to read the data, you would have to know the screen resolution to know how many bytes are needed to get the image. each pixel is 2 bytes, so if the screen res was 480x800, you would have to read 768,000 bytes for the image, since a 480x800 RGB565 image has 384,000 pixels.
For newer Android platforms, one can execute a system utility screencap in /system/bin to get the screenshot without root permission.
You can try /system/bin/screencap -h to see how to use it under adb or any shell.
By the way, I think this method is only good for single snapshot.
If we want to capture multiple frames for screen play, it will be too slow.
I don't know if there exists any other approach for a faster screen capture.
[Based on Android source code:]
At the C++ side, the SurfaceFlinger implements the captureScreen API. This is exposed over the binder IPC interface, returning each time a new ashmem area that contains the raw pixels from the screen. The actual screenshot is taken through OpenGL.
For the system C++ clients, the interface is exposed through the ScreenshotClient class, defined in <surfaceflinger_client/SurfaceComposerClient.h> for Android < 4.1; for Android > 4.1 use <gui/SurfaceComposerClient.h>
Before JB, to take a screenshot in a C++ program, this was enough:
ScreenshotClient ssc;
ssc.update();
With JB and multiple displays, it becomes slightly more complicated:
ssc.update(
android::SurfaceComposerClient::getBuiltInDisplay(
android::ISurfaceComposer::eDisplayIdMain));
Then you can access it:
do_something_with_raw_bits(ssc.getPixels(), ssc.getSize(), ...);
Using the Android source code, you can compile your own shared library to access that API, and then expose it through JNI to Java. To create a screen shot form your app, the app has to have the READ_FRAME_BUFFER permission.
But even then, apparently you can create screen shots only from system applications, i.e. ones that are signed with the same key as the system. (This part I still don't quite understand, since I'm not familiar enough with the Android Permissions system.)
Here is a piece of code, for JB 4.1 / 4.2:
#include <utils/RefBase.h>
#include <binder/IBinder.h>
#include <binder/MemoryHeapBase.h>
#include <gui/ISurfaceComposer.h>
#include <gui/SurfaceComposerClient.h>
static void do_save(const char *filename, const void *buf, size_t size) {
int out = open(filename, O_RDWR|O_CREAT, 0666);
int len = write(out, buf, size);
printf("Wrote %d bytes to out.\n", len);
close(out);
}
int main(int ac, char **av) {
android::ScreenshotClient ssc;
const void *pixels;
size_t size;
int buffer_index;
if(ssc.update(
android::SurfaceComposerClient::getBuiltInDisplay(
android::ISurfaceComposer::eDisplayIdMain)) != NO_ERROR ){
printf("Captured: w=%d, h=%d, format=%d\n");
ssc.getWidth(), ssc.getHeight(), ssc.getFormat());
size = ssc.getSize();
do_save(av[1], pixels, size);
}
else
printf(" screen shot client Captured Failed");
return 0;
}
You can try the following library: Android Screenshot Library (ASL) enables to programmatically capture screenshots from Android devices without requirement of having root access privileges. Instead, ASL utilizes a native service running in the background, started via the Android Debug Bridge (ADB) once per device boot.
According to this link, it is possible to use ddms in the tools directory of the android sdk to take screen captures.
To do this within an application (and not during development), there are also applications to do so. But as #zed_0xff points out it certainly requires root.
Framebuffer seems the way to go, it will not always contain 2+ frames like mentioned by Ryan Conrad. In my case it contained only one. I guess it depends on the frame/display size.
I tried to read the framebuffer continuously but it seems to return for a fixed amount of bytes read. In my case that is (3 410 432) bytes, which is enough to store a display frame of 854*480 RGBA (3 279 360 bytes). Yes, the frame in binary outputed from fb0 is RGBA in my device. This will most likely depend from device to device. This will be important for you to decode it =)
In my device /dev/graphics/fb0 permissions are so that only root and users from group graphics can read the fb0. graphics is a restricted group so you will probably only access fb0 with a rooted phone using su command.
Android apps have the user id (uid) app_## and group id (guid) app_## .
adb shell has uid shell and guid shell, which has much more permissions than an app.
You can actually check those permissions at /system/permissions/platform.xml
This means you will be able to read fb0 in the adb shell without root but you will not read it within the app without root.
Also, giving READ_FRAME_BUFFER and/or ACCESS_SURFACE_FLINGER permissions on AndroidManifest.xml will do nothing for a regular app because these will only work for 'signature' apps.
if you want to do screen capture from Java code in Android app AFAIK you must have Root provileges.

Trying to set a texture from Java plugin in Unity 4.6

What i'm trying to do:
Pull an image from sd-card on phone using Java Plugin.
Unity passes a texture ID to plugin.
Plugin uses opengl to assign the image to the texture in Unity through the ID.
Will (eventually) be used to play a video clip from the phone in Unity, for now, it's just trying to change a texture outside of unity.
My issue:
When i call the method in the plugin, passing texture.GetNativeTextureID() into it, the texture does not change. I'm currently only using a simple black 50x50 texture for testing, and the original texture is a flat white.
I'm worried that i've missed something significant, as this is my first time working with Gl calls in java. Much of the answers to similar problems involve using native C++ instead of Java, but I can't find a concrete answer saying that C++ must be used. I'd like to do my best to avoid writing another set of plugins and plugin handlers for C++, but if it's the most efficient/only way to get this working, i'll do it as i'm not unfamiliar with OpenGL and C++
Code:
The plugin method is called from OnPreRender() in a script attached to the main camera:
if (grabTex) {
int texPtr = m_VideoTex.GetNativeTextureID();
Debug.Log( "texPtr = " + texPtr );
m_JVInterface.SetTex( texPtr );
}
m_VideoTex is a basic Texture2D( 50, 50 ) with all pixels set to white, attached to the diffuse shader on the quad in the scene.
The Java plugin code is as follows:
public void SetTexture(Context cont, int _texPointer) {
if (_texPointer != 0) {
final BitmapFactory.Options options = new BitmapFactory.Options();
options.inScaled = false;
options.inJustDecodeBounds = false;
final Bitmap bitmap = BitmapFactory.decodeFile("/storage/emulated/0/Pictures/black.jpg", options);
GLES20.glBindTexture(GLES20.GL_TEXTURE_2D, _texPointer);
GLUtils.texImage2D(GLES20.GL_TEXTURE_2D, 0, bitmap, 0);
Log.i("VideoHandler", "Recieved ID: " + _texPointer);
bitmap.recycle();
}
}
This is most likely a problem with the OpenGL Context. The easiest way would be to send the texture as raw bytes to Unity and then upload as texture inside Unity.

Categories

Resources