Inspired by "Hipsterize Your Dog With Deep Learning"
Project contains:
- Cat face detection with pretrained Mobilenetv2
- Cat facial landmarks detection with pretrained Mobilenetv2
At first, I used cat frontal face detector of OpenCV, but it looks so bad performance for most of real cat photos. So I decided up to make new model with deep learning.
Regression method is used for both face detection and landmark detection, so that model is very naive to use on real application. But it works extremely well than I expected ;)
Used Cat dataset on Kaggle for training and validation.
- Input (Full image 224x224) - Face detection model - Output (face bounding box)
- Input (Face image 224x224) - Facial landmarks model - Output (9 landmarks points)
- Python
- Keras
- Numpy
- Dlib
- OpenCV
- Pandas
python preprocess.py
python train.py
python preprocess_lmks.py
python train_lmks.py
python test.py bbs_1.h5 lmks_1.h5
- Detect one cat per frame
- Powerful for frontal faces (a bit low performance for side faces)
- Cannot detect existence, this model thinks cat must be in the picture
- Multiple cats detection
- Data augmentation (flip, translation, rotation, noise...)
- YOLO like model (class probability map)
- Use transpose convolution layers for landmarks reconstruction (to preserve spatial information)
- Mobile implementation
- Train Dlib shape predictor model
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent