sengsational's profile - activity

Although I have not run timings, I suspect that going through formats such as XML or JSON would require more space and computing resources than Bitmap, which , as you have noted, is easily exchangable with the Mat object. Also, going though JSON requires an additional library.

Given the ability to convert to/from Mat and Bitmap, and also given the ability to convert to/from Bitmap to byte[], you have the ability to save in Android's SQLite database as a blob. The utility class below demonstrates the latter conversion:

public class DbBitmapUtility {

    // convert from bitmap to byte array
    public static byte[] getBytes(Bitmap bitmap) {
        ByteArrayOutputStream stream = new ByteArrayOutputStream();
        bitmap.compress(CompressFormat.PNG, 0, stream);
        return stream.toByteArray();
    }

    // convert from byte array to bitmap
    public static Bitmap getImage(byte[] image) {
        return BitmapFactory.decodeByteArray(image, 0, image.length);
    }
}

XML and JSON have their uses, but it would seem that in this case, since we have the ability to store and retrieve a binary without introducing an interstitial format, it might be wise to avoid their use in this case.

SQLite is quite straight forward to use:

public void insert( String name, byte[] imageByteArray) throws SQLiteException{
    SQLiteDatabase database = this.getWritableDatabase();
    ContentValues cv = new  ContentValues();
    cv.put(KEY_NAME,    name);
    cv.put(KEY_IMAGE,   imageByteArray);
    database.insert( DB_TABLE, null, cv );
}

public byte[] getByteArray(Cursor cursor) { 
    return = cursor.getBlob(BLOB_COL_NUM);  
}

@Bob Woodley, Did you ever find a way to match better?

@panc, Did you get the project working?

There is a similar question here with code on github.

There is a feature of Android Studio to Include C++ Support which generates an external build file called CMakeLists.txt. The similar question has changes required to that file to get the C++ portion of OpenCV to compile in Android Studio.

Possible duplicate questions: http://answers.opencv.org/question/90... and http://answers.opencv.org/question/99...

@Teju1106, Did you have any results with your project?

There is a similar question here with code on github.

There is a feature of Android Studio to Include C++ Support which generates an external build file called CMakeLists.txt. The similar question has changes required to that file to get the C++ portion of OpenCV to compile in Android Studio.

Possible duplicate question: http://answers.opencv.org/question/99...

@siline, why is the result matrix the difference between the sizes of the two input images?

There is a similar question here with code on github.

There is a feature of Android Studio to Include C++ Support which generates an external build file called CMakeLists.txt. The similar question has changes required to that file to get the C++ portion of OpenCV to compile in Android Studio.

Attempted to improve the question based on your inquiry. The goal is to match retail packaging in the library with one arriving on the device, so if you can improve the question, please do.

Given a set of packaging matches that numbered few enough to be completely downloaded to the Android device, what would be a good approach for detection / classification? My first thought was keypoints, since typical retail packaging has plenty of structure to work with. False positives might need attention, perhaps managed with geometric verification. HOG-template with sliding window and multiple scales as primary detection might be too process intensive for the device.

If someone here has experience in this area and knows what works and what should be avoided, that would save a lot of experimentation. Or if there's code somewhere that does something like this that works, that starting point would also be very helpful.

Update May 2017:

After putting this project aside for almost a year, I've decided to try again. Unfortunately there's way more questions here than answers.

Given a library of 500 or 1000 pictures of items from a grocery store shelf, and the requirement to detect any of those items using an Android device's camera, what would be some logical ways to construct this kind of detection app?

Although new pictures would be getting added every so-often, let's presume that the set of pictures in the library would be fixed. Thus would the data resulting from training reside on the Android device, and not depend upon support from a web site?

Neither the library pictures, nor the device user's aim will be perfect, but we can presume both are upright within a few degrees, and positioned generally in front of the object. So a rectangular object should appear roughly rectangular and roughly upright in the library and as the user positions the camera. The library image would only include only the product packaging (nothing surrounding it). The majority of the user's image will contain the product packaging. In other words, the requirement is to have the user frame a single product (although there will be artifacts on the edges).

Are there any open source projects that are similar to what is described in this question? Are there any open source projects (Android or not) that load training data from a set of images into blobs in a database, then compare those blobs to the "current" image?

Would a typical Android device have enough processing power to check video frames, or would the user need to "take a picture" and then let the app churn to get a result?

On this page concerning installation, it currently says:

You can choose downloading ADT Bundle package that in addition to Android SDK Tools includes Eclipse + ADT + NDK/CDT plugins, Android Platform-tools, the latest Android platform and the latest Android system image for the emulator - this is the best choice for those who is setting up Android development environment the first time!

But this is no longer a supported development environment, as indicated here.

How do required changes to the OpenCV documentation get in the queue?