AOF-Face-Recognition

I would like to launch a piece of research on facial recognition that especially serves the AOF community. I will learn about the underlying mechanism of facial recognition (and machine learning algorithms that drives it) and the tools that are used to implement it. I will then try to train a model with pictures of members of the AOF community and write codes to recognize people in media including photos and videos. This model and tools I produce would be helpful for future projects like, for example, even more convenient check-in for school events by scanning faces. In addition, my code and model (but not any actual data from images of the school community) will be open-sourced on GitHub. Other people and their projects can benefit from what I have created. It would encourage our students to use what they have learned in class to do great contributions to the community.

How to use it

All script should be run at the root directory of this project.

Run script under the demo folder. Now the system can recognized students enrolled during 2015-2019.

###To train new models:

1. Remove all files under the data folder but keep the folders themselves.
2. Put photos under training_data/known. Change the function transform_yearbook_photos in util/general.py to convert filename into actual name displayed.
3. Change settings on top in processor.py and processor_num_map.py if needed. Run processor_num_map.py to generate encoding data for training examples. These information will be stored to data/cache/preprocess
4. Use add function in test_data/test_manager.py to add at least one test set. Change parameters on top if needed.
5. Choose a model to use. In the folder, run xxx_test.py to generate the model and test it. Model files will be generated in data/model/xxx
6. Create a EncodingsPredictor of that model. Set the correct parameters. Now you can use it with demo or with your own code.

###Testing: There are tests for each model as well as time_test.py and all_test.py to test multiple models.

###Model Suggestions: Usage

• Universal: NN
• Low Res: NN
• Prevent False Positive: NN
• Prevent False Negative: KNN
• Tens of thousands of training data: NN or KNNKmeans

Training

• Mode recognition speed: NN < KNNKmeans < KNN
• Mode recognition speed (huge amount of training data): NN < KNNKmeans << KNN
• Model accuracy: NN = KNN > KNNKmeans
• Model accuracy (huge amount of training data): NN = KNN = KNNKmeans
• Model training speed and hardware requirement: KNN < KNNKmeans << NN

How well do the current model do: