DeepGenomotion is an android application that uses deep learning inspired augemented reality filters.
We use inference from our custom gender and emotion model to draw ar filter on real-time camera feed with inference rate of 25
to 100
ms for inference on single image.
- Gender Model is a
MobileNet_V1_224_0.5
model trained on IMDB-WIKI celebrity face dataset. - Emotion Model is a
MobileNet_V2_224_1.4
model trained on AffectNet Dataset containing 1m facial images and 11 classes of emotion - Both model are trained using tensorflow and then the trained graph is converted into Tensorflow Lite for serving.
- We use MobileVision API for facial landmark detection.