Pex-tech-challenge

Goal

To create a dataset of indoor and outdoor images
To train a model to identify images as belonging to the indoor or outdoor class
To build a CLI tool to input images one by one, and for the tool to predict the class that the image belongs to

Methodology

I used the script from this codebase to extract relevant videos from the YouTube-8m dataset. The categories I queried for indoor images were Room, Bedroom, Classroom, Office, Gym, Couch, Cooking, and Chair. The categories I queried for outdoor images were Beach, Landscape, Skyscraper, Mountain, Hiking, Skiing, Weather, and Thunderstorm. After obtaining these videos, I extracted 5 frames from each video to create a dataset of images.

I used a pretrained VGG16 model, excluding the fully connected layers. Instead, I added fully connected layers of my own, and trained only those layers on the dataset that I had collected. Due to the lack of compute power on my local machine, I trained it only for 5 epochs. I was able to obtain a validation accuracy of 79.92%.

I then created a command line tool called "mytool", which would take an image path as an argument, and would output whether the image was taken indoors or outdoors.

Running the code

Install python 2.7
Install virtualenv
Clone this repository
cd into this repository
Create a virtual environment
Run pip install -r requirements.txt
Run python setup.py install
Run mytool --img img_name where img_name is the path to an image of your choice

Important notes

The main source code is contained in the 'src' directory

utils.py was used to create the dataset. It was run as python utils.py indoor or python utils.py outdoor depending upon the data I was creating.

model.py contains the code for training the deep learning model. It took no other arguments, and simply ran with 'python model.py'

unit_test.py contains a few test units that I wrote

All the data related content is contained in the 'data' directory

It contains a directory titled photos, which contains the dataset I created.

It was in this folder that I had two other folders, each containing the indoor and outdoor videos respectively. Due to the size of the videos, I did not upload them onto github.

It also has two bash scripts, which were used to make the data processing more convenient.

abhishekvasu94 / pex-tech-challenge Goto Github PK