Develop Reference

Image Normalization Prior to Pytorch_Mobile Prediction in Dart/Flutter?

I am attempting to run an image through a pytorch model I have created and incorporated into a mobile flutter app. In the Training process, prior to moving the model to my mobile dev environment, I normalize my training images in python using Pytorch Transforms with the following line of code:
transforms.Normalize(mean=[0.75107294, 0.51543763, 0.52209598], std=[0.13829332, 0.15216838, 0.16517265]),
I have tried to figure this out from the documentation for the Pytorch_Mobile package for Dart/Flutter but I am unsure as to wether I have to normalize my image prior to passing it to the network for a prediction, my code currently looks as follows (filePath being the Path to the Image):
File resizedFile = await FlutterNativeImage.compressImage(filePath,
quality: 100, targetWidth: 224, targetHeight: 224);
List? classificationPrediction =
await classificationModel.getImagePredictionList(resizedFile, 224, 224);
However, when passing images to the model through my flutter app I am receiving results that just don't seem to reflect the accuracy that was achieved through the test dataset used in Training in Pytorch in Python.
Could the issue be that I need to normalize the image prior to passing it to the model in Flutter, and if so, how would I best achieve this?

Generate and export point cloud from Project Tango

After some weeks of waiting I finally have my Project Tango. My idea is to create an app that generates a point cloud of my room and exports this to .xyz data. I'll then use the .xyz file to show the point cloud in a browser! I started off by compiling and adjusting the point cloud example that's on Google's github.
Right now I use the onXyzIjAvailable(TangoXyzIjData tangoXyzIjData) to get a frame of x y and z values; the points. I then save these frames in a PCLManager in the form of Vector3. After I'm done scanning my room, I simple write all the Vector3 from the PCLManager to a .xyz file using:
OutputStream os = new FileOutputStream(file);
size = pointCloud.size();
for (int i = 0; i < size; i++) {
String row = String.valueOf(pointCloud.get(i).x) + " "
+ String.valueOf(pointCloud.get(i).y) + " "
+ String.valueOf(pointCloud.get(i).z) + "\r\n";
os.write(row.getBytes());
}
os.close();
Everything works fine, not compilation errors or crashes. The only thing that seems to be going wrong is the rotation or translation of the points in the cloud. When I view the point cloud everything is messed up; the area I scanned is not recognizable, though the amount of points is the same as recorded.
Could this have to do something with the fact that I don't use PoseData together with the XyzIjData? I'm kind of new to this subject and have a hard time understanding what the PoseData exactly does. Could someone explain it to me and help me fix my point cloud?

Yes, you have to use TangoPoseData.
I guess you are using TangoXyzIjData correctly; but the data you get this way is relative to where the device is and how the device is tilted when you take the shot.
Here's how i solved this:
I started from java_point_to_point_example. In this example they get the coords of 2 different points with 2 different coordinate system and then write those coordinates wrt the base Coordinate frame pair.
First of all you have to setup your exstrinsics, so you'll be able to perform all the transformations you'll need. To do that I call mExstrinsics = setupExtrinsics(mTango) function at the end of my setTangoListener() function. Here's the code (that you can find also in the example I linked above).
private DeviceExtrinsics setupExtrinsics(Tango mTango) {
//camera to IMU tranform
TangoCoordinateFramePair framePair = new TangoCoordinateFramePair();
framePair.baseFrame = TangoPoseData.COORDINATE_FRAME_IMU;
framePair.targetFrame = TangoPoseData.COORDINATE_FRAME_CAMERA_COLOR;
TangoPoseData imu_T_rgb = mTango.getPoseAtTime(0.0,framePair);
//IMU to device transform
framePair.targetFrame = TangoPoseData.COORDINATE_FRAME_DEVICE;
TangoPoseData imu_T_device = mTango.getPoseAtTime(0.0,framePair);
//IMU to depth transform
framePair.targetFrame = TangoPoseData.COORDINATE_FRAME_CAMERA_DEPTH;
TangoPoseData imu_T_depth = mTango.getPoseAtTime(0.0,framePair);
return new DeviceExtrinsics(imu_T_device,imu_T_rgb,imu_T_depth);
}
Then when you get the point Cloud you have to "normalize" it. Using your exstrinsics is pretty simple:
public ArrayList<Vector3> normalize(TangoXyzIjData cloud, TangoPoseData cameraPose, DeviceExtrinsics extrinsics) {
ArrayList<Vector3> normalizedCloud = new ArrayList<>();
TangoPoseData camera_T_imu = ScenePoseCalculator.matrixToTangoPose(extrinsics.getDeviceTDepthCamera());
while (cloud.xyz.hasRemaining()) {
Vector3 rotatedV = ScenePoseCalculator.getPointInEngineFrame(
new Vector3(cloud.xyz.get(),cloud.xyz.get(),cloud.xyz.get()),
camera_T_imu,
cameraPose
);
normalizedCloud.add(rotatedV);
}
return normalizedCloud;
}
This should be enough, now you have a point cloud wrt you base frame of reference.
If you overimpose two or more of this "normalized" cloud you can get the 3D representation of your room.
There is another way to do this with rotation matrix, explained here.
My solution is pretty slow (it takes around 700ms to the dev kit to normalize a cloud of ~3000 points), so it is not suitable for a real time application for 3D reconstruction.
Atm i'm trying to use Tango 3D Reconstruction Library in C using NDK and JNI. The library is well documented but it is very painful to set up your environment and start using JNI. (I'm stuck at the moment in fact).
Drifting
There still is a problem when I turn around with the device. It seems that the point cloud spreads out a lot.
I guess you are experiencing some drifting.
Drifting happens when you use Motion Tracking alone: it consist of a lot of very small error in estimating your Pose that all together cause a big error in your pose relative to the world. For instance if you take your tango device and you walk in a circle tracking your TangoPoseData and then you draw you trajectory in a spreadsheet or whatever you want you'll notice that the Tablet will never return at his starting point because he is drifting away.
Solution to that is using Area Learning.
If you have no clear ideas about this topic i'll suggest watching this talk from Google I/O 2016. It will cover lots of point and give you a nice introduction.
Using area learning is quite simple.
You have just to change your base frame of reference in TangoPoseData.COORDINATE_FRAME_AREA_DESCRIPTION. In this way you tell your Tango to estimate his pose not wrt on where it was when you launched the app but wrt some fixed point in the area.
Here's my code:
private static final ArrayList<TangoCoordinateFramePair> FRAME_PAIRS =
new ArrayList<TangoCoordinateFramePair>();
{
FRAME_PAIRS.add(new TangoCoordinateFramePair(
TangoPoseData.COORDINATE_FRAME_AREA_DESCRIPTION,
TangoPoseData.COORDINATE_FRAME_DEVICE
));
}
Now you can use this FRAME_PAIRS as usual.
Then you have to modify your TangoConfig in order to issue Tango to use Area Learning using the key TangoConfig.KEY_BOOLEAN_DRIFT_CORRECTION. Remember that when using TangoConfig.KEY_BOOLEAN_DRIFT_CORRECTION you CAN'T use learningmode and load ADF (area description file).
So you cant use:
TangoConfig.KEY_BOOLEAN_LEARNINGMODE
TangoConfig.KEY_STRING_AREADESCRIPTION
Here's how I initialize TangoConfig in my app:
TangoConfig config = tango.getConfig(TangoConfig.CONFIG_TYPE_DEFAULT);
//Turning depth sensor on.
config.putBoolean(TangoConfig.KEY_BOOLEAN_DEPTH, true);
//Turning motiontracking on.
config.putBoolean(TangoConfig.KEY_BOOLEAN_MOTIONTRACKING,true);
//If tango gets stuck he tries to autorecover himself.
config.putBoolean(TangoConfig.KEY_BOOLEAN_AUTORECOVERY,true);
//Tango tries to store and remember places and rooms,
//this is used to reduce drifting.
config.putBoolean(TangoConfig.KEY_BOOLEAN_DRIFT_CORRECTION,true);
//Turns the color camera on.
config.putBoolean(TangoConfig.KEY_BOOLEAN_COLORCAMERA, true);
Using this technique you'll get rid of those spreads.
PS
In the Talk i linked above, at around 22:35 they show you how to port your application to Area Learning. In their example they use TangoConfig.KEY_BOOLEAN_ENABLE_DRIFT_CORRECTION. This key does not exist anymore (at least in Java API). Use TangoConfig.KEY_BOOLEAN_DRIFT_CORRECTION instead.

Minko - Android light issue

I am working with minko and seem to be facing a light issue with Android.
I managed to compile for linux64, Android and html a modified code (based on the tutorials provided by Minko). I simply load and rotate 4 .obj files (the pirate one provided and 3 found on turbosquid for demo purposes only).
The correct result is viewed in the linux64 and html version but the Android one has a "redish" light thrown into it, although the binaries are being generated from the same c++ code.
Here are some pics to demonstrate the problem:
linux64 :
http://tinypic.com/r/qzm2s5/8
Android version :
http://tinypic.com/r/23mn0p3/8
(Couldn’t link the html version but it is close to the linux64 one.)
Here is the part of the code related to the light :
// create the spot light node
auto spotLightNode = scene::Node::create("spotLight");
// change the spot light position
//spotLightNode->addComponent(Transform::create(Matrix4x4::create()->lookAt(Vector3::zero(), Vector3::create(0.1f, 2.f, 0.f)))); //ok linux - html
spotLightNode->addComponent(Transform::create(Matrix4x4::create()->lookAt(Vector3::zero(), Vector3::create(0.1f, 8.f, 0.f))));
// create the point light component
auto spotLight = SpotLight::create(.15f, .4f); //ok linux and html
// update the spot light component attributes
spotLight->diffuse(4.5f); //ori - ok linux - html
// add the component to the spot light node
spotLightNode->addComponent(spotLight);
//sets a red color to our spot light
//spotLightNode->component<SpotLight>()->color()->setTo(2.0f, 1.0f, 1.0f);
// add the node to the root of the scene graph
rootNode->addChild(spotLightNode);
As you can notice the color()->setTo has been turned off and works for all except Android (clean and rebuild). Any idea what might be the source of the problem here ?
Any pointer would be much appreciated.
Thx.

Can you test it on other Android devices or with a more recent ROM and give us the result? LG-D855 (LG G3) is powered by an Adreno 330: those GPUs are known to have GLSL compiling deffects, especially with loops and/or structs like we use in Phong.fragment.glsl on the master branch.
The Phong.fragment.glsl on the dev branch has been heavily refactored to fix this (for directional lights only for now).
You could try the dev branch and a directional light and see if it fixes the issue. Be careful though: the dev branch introduces the beta 3, with some API changes. The biggest API change being the math API now using GLM, and the *.effect file format. The best way to go is simply to update your math code to use the new API, everything else should be straight forward.

connectOnFrameAvailable() provides TangoImageBuffer with curious format infos

Also trying to get access to color data bytes from color cam of Tango, I was stuck on java API by being able to connect tango Cam to a surface for display (but just OK for display in fact, no easy access to raw data, nor time stamp)... so finally I switch using C API on native code (latest FERMAT lib and header) and follow recommendation I found on stack Overflow by registering a derivated sample code to connectOnFrameAvailable()... (I start using PointCloudActivity sample for that test).
First problem I found is somewhat a side effect of registering to that callback, that works usually fine (callbacks gets fire regularly), but then another callback that I also registered, to get xyz clouds, start to fail to fire. Like in sample code I mentioned, clouds are get through a onXYZijAvailable() callback, that the app registers using TangoService_connectOnXYZijAvailable(onXYZijAvailable).
So failing to get xyz callback fired is not happening always, but usually half of the time, during tests, with a awful workaround that is by taking the app in background then foreground again ... this is curious, is this "recover" related to On-pause/On-resume low level stuff??). If someone has clues ....
By the way in Java API, same side effect was observed, once connecting cam texture for display (through Tango adequate API ...)
But here is my second "problem", back to acquiring YV12 color data from camera :
through registering to TangoService_connectOnFrameAvailable( TangoCameraId::TANGO_CAMERA_COLOR, nullptr, onFrameAvailable)
and providing static funtion onFrameAvailable defined like this :
static void onFrameAvailable(void* ctx, TangoCameraId id, const TangoImageBuffer* buffer)
{
...
LOGI("OnFrameAvailable(): Cam frame data received");
// Check if data format of expected type : YV12 , i.e.
// TangoImageFormatType::TANGO_HAL_PIXEL_FORMAT_YV12
// i.e. = 0x32315659 // YCrCb 4:2:0 Planar
//LOGI("OnFrameAvailable(): Frame data format (%x)", buffer->format);
....
}
the problem is that width, height, stride information of received TangoImageBuffer structure seems valid (1280x720, ...), BUT the format returned is changing every-time, and not the expected magic number (here 0x32315659) ...
I am doing something wrong there ? (but other info are OK ...)
Also, there is apparently only one data format defined (YV12 ) here, but seeing Fish Eye images from demo app, it seems grey level image, is it using same (color) format as low level capture than the RGB cam ???

1) Regarding the image from the camera, I came to the same conclusion you did - only availability of image data is through the C API
2) Regarding the image - I haven't had any issues with YUV, and my last encounter with this stuff was when I wrote JPEG stuff - the format is naked, i.e. it's an organizational structure and has no header information save the undefined metadata in the first line of pixels mentioned here - Here's a link to some code that may help you decode the image in a response to another message here
3) Regarding point cloud returns -
Please note this information is anecdotal, and to some degree the product of superstition - what works for me only does that sometimes, and may not work at all for you
Tango does seem to have a remarkable knack to simply stop producing point clouds. I think a lot of it has to do with very sensitive timing internally (I wonder if anyone mentioned that Linux ain't an RTOS when this was first crafted)
Almost all issues I encounter can be attributed to screwing up the timing where
A. Debugging the C level can may point clouds stop coming
B. Bugs in the native or java code that cause hiccups in the threads that are handling the callbacks can cause point clouds to stop coming
C. Excessive load can cause the system to loose sync, at which point the point clouds will stop coming - this is detectable, you will start to see a silvery grid pattern appear in rectangular areas of the image, and point clouds will cease. Rarely, the system will recover if load decreases, the silvery pattern goes away, and point clouds come back - more commonly the silvery pattern (I think its the 3d spatializing grid) grows to cover more of the image - at least a restart of the app is required for me, and a full tablet reboot every 3rd time or so
Summarizing, that's my suspicions and countermeasures, but it's based completely on personal experience -

How to get pixel data from screencap.cpp directly

I'm a newbie in Android.
I using Nexus7 reference device and I've downloaded the full source code from source.android.com.
I have an engineering system image and I can make a system application.
/system/bin/screencap utility is good for me to capture screen.
I want to get a pixel data using screencap.cpp directly in my application.
When I used to screencap utility, the process is like below.
capture screen and save an image.
Open image file
decodefile to bitmap
get pixel data(int array) from bitmap
I want to remove the step 1, 2 and 3.
Just call api to get pixel data of a screen directly,
How can I do that?

If you're running with system privileges, you can just ask SurfaceComposerClient for the pixel data rather than launching a separate process to do it for you.
Looking at the screencap source code, all you really need is the Binder initialization:
ProcessState::self()->startThreadPool();
and the SurfaceComposerClient IPC call:
ScreenshotClient screenshot;
sp<IBinder> display = SurfaceComposerClient::getBuiltInDisplay(displayId);
if (display != NULL && screenshot.update(display, Rect(), false) == NO_ERROR) {
base = screenshot.getPixels();
w = screenshot.getWidth();
h = screenshot.getHeight();
s = screenshot.getStride();
f = screenshot.getFormat();
size = screenshot.getSize();
}
You can safely ignore all the /dev/graphics/fb0 stuff below it -- it's an older approach that no longer works. The rest of the code is just needed for the PNG compression.
If you're not running with system privileges, you can't capture the entire screen. You can capture your app though.

If you are writing a Java app just call /system/bin/screencap from your application (using java.lang.Process) and read the result into memory as a binary stream. You can see the binary structure in screencap.cpp, but it's just width, height, and format as four byte integers followed by the data.
Note to other readers: this is only be possible if you are a system app.

1) You can transfer data from your screencap utility to your App over the network by using sockets.
2) Android NDK can be used for direct function calls of your utility from your App.