- Trung Vu
- Sihyun Lee
- Alexandre Lamy
This is the Github repo for the Denoising Face Recognition via DBSCAN project that we worked on for the class COMS 4995: Methods in Unsupervised Learning (Fall 2018). It contains the code and data that we used as well as the results we obtained. It also has our final poster and writeup.
The poster is the file final_poster.pdf, the final report is the file final_report/final_report.pdf. Results can all be found in the results/ folder. Here is a short table describing each of the
videos:
| Video File | Description |
|---|---|
| BIGBANG | Full original video |
| BIGBANG_ANN | Full original video with detection/recognition bounding boxes |
| BIGBANG_DBSCAN | Full original video with cleaned bounding boxes (after post processing by our algorithm) |
| BIGBANG_CROP | Most interesting portion of full video |
| BIGBANG_CROP_ANN | Most interesting portion of full video with detection/recognition bounding boxes |
| BIGBANG_CROP_DBSCAN | Most interesting portion of full video with cleaned bounding boxes (after post processing by our algorithm) |
To clone this repo first make sure that git lfs is installed (it helps store large files). If not run the following commands.
wget https://github.com/git-lfs/git-lfs/releases/download/v2.6.0/git-lfs-linux-amd64-v2.6.0.tar.gz
tar -xf git-lfs-linux-amd64-v2.6.0.tar.gz
sudo ./install.sh
Once this is done you can git clone, git pull, git add, and git commit as usual.
Our raw data (the images on which the neural net was trained, the neural net parameters, etc.)
are all in the data/ folder. However, due to the very large number of files in this folder we do NOT
store it directly on Github. Instead, we zip it and store the data.zip file using git lfs. To
retrieve the data/ folder, first make sure your repo is up to date (do a git pull) and then simply
run
make getdata
this will automatically unzip data.zip and create the data/ folder.
For the code to work, please make sure you are running python 3.6 and install the python packages
numpy, scikit-learn, and opencv-python.
Almost all our code is broken down into independent and self contained scripts that take in command line arguments. Each can be called with the -h
flag to see details.
The parameters for the Neural Network which creates the image embedding is found in file data/vgg2.pb. These weights come from the vgg2 dataset (we did not retrain).
We used standard facenet scripts to train an SVM classifier over the 5 faces we were interested in detecting. These scripts are not included but are easy to find online (see facenet).
The resulting SVM model is included in code/classifier.pkl. The file code/classify_video.py is a simple script that uses that classifier to detect and classify faces in a video.
The code for our post processing method can then be found in the file code/dbscan_denoising_alg.py.
If you make changes to the data (try new videos or people, change model weight, etc.) there is a simple procedure to push those changes to Github.
Simply use the following command to automatically correspondingly update the data.zip file.
make zipdata
You can then git add, git commit, and git push as usual.