Task: Build a simple Web Application, which when given an image, is able to detect a face (show with a boundary box) and give a result saying whether Narendra Modi and/or Arvind Kejriwal are present in the image or not.

Brief description:

1. Used 3 different classifiers to process the image - A Haar Cascade clf to detect and draw boundry around the face and 2 Convolutional Neural Networks to further evaluate if it is Arvind Kejriwal and/or Narendra Modi.
2. Key Challenge - Sampling of images (positive and negative sets) to train CNNs. Segregation & and keeping negative samples of each kind in the cross-positive categories to build an effective classification system. In addition, keeping the training & validation datasets mutually exclusive. Also used flow_generator to overcome the limitation of limited data to create multiple variants of the same image.

Particulars:

• Data Collection: Used Fatkun Batch Image chrome extension to extract data from Google Images.
• Train-Validation split: Nearly 80:20 split. 80% data used to train, the rest is used to cross-validate.
• Training carried out on FloydHub's CPU instances.

Given higher compute, the same/different architectures can be experimented with higher number of epochs

• Deployment: DigitalOcean droplet.

Directory Structure:

.
.
├── collections                     # MongoDB exports of data in GridFS specification
├── src                             # source Files
│   ├── train.py                    # training script
│   ├── dump_collections.py         # script used to dump data into GridFS specs.
│   ├── extract_img_example.py      # to extract data from gridFS specs
├── models                          # saved models               
├── static
├── templates
├── precog_task.py                  # execution script  
├── utils.py
└── wsgi.py
.
.

Structure of collection:

The data is stored in GridFS specs in MongoDB. dump_collections.py was used to export data to MongoDB. To pass off all the data into the db, the script can be executed as - python3 src/dump_collections.py -label "negative" -part "train" -path "/Users/sominwadhwa/Desktop/temp/nm/train/negative" . Entries stored are of the form:

{img_file_1, 'label':'positive', 'part':'train', 'person':'Arvind Kejriwal'}

Since the entire data is stored in GridFS specs, to extract it you can use the extract_img_example.py by first adding the collections to a temporary database -

sudo mongorestore --db data_for_task --collection fs.chunks /data_for_task/fs.chunks.bson
sudo mongorestore --db data_for_task --collection fs.chunks /data_for_task/fs.files.bson

To extract all images for Arvind Kejriwal, with positive class label & as part of validation dataset -

python3 extract_img_example.py -person "Arvind Kejriwal" -label "positive" -part "validation"

The entirety of this minimal-webapp has a few limitations. Due to the nature of the training data (preserve covariance, reduce randomness in samples), the detection of whether the image is of AK or NaMo will only happen if haar-cascade clf is able to detect a face, if however even a single face is detected, CNNs might be able to correctly classify the image. Altering scale_factor param of haar_cascade might as well alter the results in cases where facial propotionals are unusally large as compared to the size of the image. -

Cases where this may happen:

• Face covered with some form of obstacle.
• Unusually large face proportions as compared to the size of the photo.
• Varying viewing angles.

How to overcome this? RCNNs is an approach that can be tried out to resolve this sort of an issue.

The following samples were taken at the time of development, actual testing results may vary as the final model (weights) get updated periodically.