Project instructions are found in the Udacity README.
imread, cvColor, rectangle, resize, CascadeClassifier(...).detectMultiScale(...)
For face detection, the OpenCV function CascadeClassifier::detectMultiScale was used with an OpenCV XML file encoding a Haar Cascade classifier.
For keypoint detection, a CNN architecture was chosen, similar to that described in the NaimishNet paper. Three different variants of the architecture were tested, "full", "lite" and "superlite". The implementation of each model is found in the respective "model_xxx.py" file. A summary of the differences between the architectures, and other design decisions such as hyperparameter selection, are found in answers to three questions near the bottom of this notebook. The networks were trained on a dataset of ~3500 images labeled with truth keypoints. The amount of training was limited further by time, given the available computation resources.
This notebook contains the full pipeline: face detection, extraction and preprocessing, and keypoint prediction, as well as the final results.
After face detection, with bounding boxes drawn around the detected faces:
After keypoint prediction, using the trained "naimishlite" model:
Placement of keypoints on some of the faces is suboptimal due to limited training time and data.