This is an Android app developed for Asus Zenbo Junior Robot using embedded A.I. vision. Zenbo Junior is an AI-enabled robot platform with Android system developed by Asus, Inc. This app is able to perform gender, emotion and age detection in timely. I used android.hardware.camera2 package to invoke the system camera, and I used MobileNet as my neural network following this paper. I trained models using Tensorflow and converted them to TFLite format at the end.
-
Gender Model
- Built on IMDB - WIKI dataset which contains 500k+ celebrity faces
- The model uses
Mobile_Net_v1_224_0.5version of the MobileNet
-
Emotion Model
- Built on fer2013 dataset of Kaggle with more than 28,709 facial images in seven discrete facial expressions(happiness, anger, surprise, disgust, sadness, fear, and neutral).
- The model uses
Mobile_Net_v2_224_1.4version of the MobileNet
-
Age Model
- I followed the instruction of this paper
- Built on UTKFace dataset with 20k+ face images in the wild (only single face in one image)
- The model uses
Mobile_Net_v2_224_1.00version of the MobileNet
You may find the presentation for this project at this link. You may find the demo for this project at this link.
TODOs:
- Draw reactangles during face detection
- Optimize the age model
- New greeting after detecting different faces