Mobilenet+SSD face detector training

This repo contains code for Mobilenet+SSD face detector training. This detector is compatible with Movidius Neural Compute Stick. You need <a href="https://github.com/movidius/ncsdk" target="_blank">NCSDK</a> to test it with Neural Compute Stick.

Deploying models for Caffe and Neural Compute Stick

You can deploy two different SSD face detectors: "full" detector or "short" detector. The latter is shortened: layers 14-17 are deleted. It is a bit faster (67 ms vs 75 ms) and captures small faces only.

To deploy detectors to Caffe:

make deploy_full

or

make deploy_short

To deploy detectors to NCS (and Caffe):

make compile_full

or

make compile_short

Deployment models are placed in models/deploy.

Training

To train this detector (<a href="https://github.com/weiliu89/caffe/tree/ssd" target="_blank">SSD-Caffe</a> is needed):

1. Download <a href="http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/" target="_blank">WIDER</a> and <a href="http://vis-www.cs.umass.edu/fddb/" target="_blank">FDDB</a> datasets.

2. Edit Makefile: set data_dir, lmdb_pyscript, caffe_exec, datasets names and path to data folder.

3. Make LMBD database:

make lmdb

4. Make face model (generate templates and get pre-trained weights):

make face_model_full

5. Edit train_files/solver_train_full.prototxt if necessary and train net:

make train_full

Or resume from snapshot:

echo /path/to/snapshot > train_files/snapshot.txt
make resume_full

6. Test model:

echo /path/to/snapshot > train_files/snapshot.txt
make test_full

Test best model from this repo:

make test_best_full

7. (Optional) Make long-range (shorter) model:

make face_model_short

And test it:

make test_short_init

8. Plot loss from Caffe logs:

make plot_loss

Plot Average Precision from snapshots:

echo /path/to/any/snapshot > train_files/snapshot.txt
make plot_map_full

9. Profile initial VOC net, best face net, short face net for Neural Compute Stick:

make profile_initial

or

make profile_face_full

or

make profile_short_init

Also see Caffe_face.ipynb for details.

See images/output to see how nets perform on examples (test network to get these results).

See <a href="https://colab.research.google.com/drive/1LExcFZO8vN46xrJ8deG159eIUaW0kB-H" target="_blank">this notebook</a> for training this model in Google Colaboratory.