Skip to content

Project Changes

Jump to bottom
Shubham Panchal edited this page Jan 3, 2023 · 2 revisions

Project Changes ( Dec 2020 - Present )

Updates - September 2022

  • Modified settings.gradle to use the new plugin management system.
  • The conversion of Bitmap to NV21-formatted ByteArray ( YUV420 ) is now transformed into a suspending function to avoid blocking of the UI thread when a large number of images are being processed.

Updates - December 2021

  • Users can now control the use of GpuDelegate and XNNPack using useGpu and useXNNPack in MainActivity.kt,
 // Use the device's GPU to perform faster computations.
 // Refer https://www.tensorflow.org/lite/performance/gpu
 private val useGpu = true

 // Use XNNPack to accelerate inference.
 // Refer https://blog.tensorflow.org/2020/07/accelerating-tensorflow-lite-xnnpack-integration.html
 private val useXNNPack = true

Updates - October 2021

  • The app now has a face mask detection feature with models obtained from achen353/Face-Mask-Detector repo. You may off it by setting isMaskDetectionOn in FrameAnalyser.kt to false.

  • The source of the FaceNet model is now Sefik Ilkin Serengil's DeepFace, a lightweight framework for face recognition and facial attribute analysis. Hence, the users can now use two models, FaceNet and FaceNet512. Also, the int-8 quantized versions of these models are also available. See the following line ine MainActivity.kt,

private val modelInfo = Models.FACENET

You may use different configurations in the Models class.

  • The app will now classify users, whose images were not scanned from the images folder, as UNKNOWN. The app uses thresholds both for L2 norm and cosine similarity to achieve this functionality.

  • For requesting the CAMERA permission and access to the images folder, the request code is now handled by the system itself. See Request app permissions.

Updates - July 2021

  • We'll now use the PreviewView from Camera instead of directly using the TextureView. See the official Android documentation for PreviewView

  • As of Android 10, apps couldn't access the root of the internal storage directly. So, we've implemented Scoped Storage, where the user has to allow the app to use the contents of a particular directory. In our case, users now have to choose the images/ directory manually. See Grant access to a directory's contents.

  • The feature request #11 for serializing the image data has been considered now. The app won't load the images everytime so as to ensure a faster start.

  • The feature request #6 has also been considered. After considering the use of PreviewView, the app can now be sed in the landscape orientation.

  • The project is now backwards compatible to API level 25. For other details, see the build.gradle file.

  • The lens facing has been changed to FRONT and users won't be able to change the lens facing. The app will open the front camera of the device as a default.

  • The source of the FaceNet Keras model -> nyoki-mtl/keras-facenet

  • The image normalization step is now included in the TFLite model itself using a custom layer. We only need to cast images to float32 using the CastOp from TFLite Support Library.

  • A TextView is now added on the screen which logs important information like number of images scanned, similarity score for users, etc.

Updates - June 2021

  • The source of the FaceNet model has been changed. We'll now use the FaceNet model from sirius-ai/MobileFaceNet_TF
  • The project is now backwards compatible to API level 23 ( Android Marshmallow )
minSdkVersion 23

Updates - December 2020

  • Lens Facing of the camera can be changed now. A button is provided on the main screen itself.
  • For multiple images for a single user, we compute the score for each image. An average score is computed for each group. The group with the best score is chosen as the output. See FrameAnalyser.kt.
images ->  
    Rahul -> 
         image_rahul_1.png -> score=0.6 --- | average = 0.65 --- |
         image_rahul_2.png -> score=0.5 ----|                    | --- output -> "Rahul"
    Neeta ->                                                     |
         image_neeta_1.png -> score=0.4 --- | average = 0.35 --- |
         image_neeta_2.png -> score=0.3 ----|
  • Cosine similarity can be used alongside L2 norm. See the metricToBeUsed variable in FrameAnalyser.kt.
  • A new parameter has been added in MainActivity.kt. The cropWithBBoxes argument allows you to run the Firebase MLKit module on the images provided. If you are already providing cropped images in the images/ folder, set this argument to false. On setting the value to true, Firebase ML Kit will crop faces from the images and then run the FaceNet model on it.
Clone this wiki locally