A Deep Learning Project Produced by
Arjun Rajpal
Andrew Shao
Jai Gupta
- About Repo/Motivation
- About the Data
- Files/Dirs and what they do
- Set up and use
On the battlefield, soldiers, at times, can find it difficult to identify potential targets in the fog of war. It can be hard to spot a tank hiding in a well-placed bush or an aircraft strafing through heavy smog, leading to surprise attacks and potential casualties. Our project will examine ways to make near real-time identification of aircraft with various data augmentation techniques to address this difficulty potentially. In short, we want to find and evaluate different ways of performing near real-time object detection based on both speed of detection and accuracy.
-
Military Aircraft Dataset (https://www.kaggle.com/datasets/a2015003713/militaryaircraftdetectiondataset)
a. This dataset contains images for 43 different types of military aircraft from various angles and locations. This includes bounding given in PASCAL VOC format to identify where the objects lie within each image. These images are sourced from Google. This image dataset meets our requirements because it contains images containing military aircraft. Moreover, we will be doing substantial augmentation to expand this dataset. In total, there are 19746 images in the dataset spread across 43 different classes. This distribution of images by class is given in the image below.
\eda.ipynb - python notebook for all of our initial eda on the military aircraft dataset
\eda_model_trainCNNs.ipynb - python notebook to do more eda + train our Resnet, VGG and Alex Net Models
\evaluation_on_original.ipynb - evaluation code for our model (class accuracy/precision)
\visualization.ipynb - overall model performance visualizations
\FasterRCNNModel_training.ipynb - training file for Faster RCNN model
\video_processing.ipynb - file to evaluate all of our models on the video analysis aspect
\videos - a place to store desired videos to be processed - video file auto downloads so expect this to be empty upon cloning this repo
\models - a place to store models which are finished training **MUST Create**
Special Note on Directories: Our models themselves were too large to upload to GitHub, so one must create a models folder once the repo is cloned + train models using the eda_model_trainCNNS.ipynb and FasterRCNNModel files
unzip models.zip
Likewise, since the dataset was too large to upload, one must download from the link about and extract to a cloned version of the repo (you know you have done it right when the file folder archive is present)
Note: Since we were using colab a lot of the commands that were used to install libraries are included in the notebooks
- First, download data from the link
- Create a models folder
- Run desired files
This repo has code that depends on hardware (GPUs), and it is recommended that one uses this repo in a Google Colab-type environment as that is where these files were developed.