High-Level Project Summary
We develop an application to help detect military vehicles on the battlefield using drones and satellites. It solves the challenge in a way, that neural networks could process pictures faster than a human beings. It is important for Ukrainian soldiers to detect the russian army in time.
Link to Final Project
Link to Project "Demo"
Detailed Project Description
To detect armored vehicles on drone images - we decided to train the existing model YOLOv5 on images from drones. Unfortunately, we could not find a large dataset containing such images. We then have taken the small dataset with videos from drones, that were collected by UCU https://storage.googleapis.com/drone_vehicle_footage_dataset_public.
As the model cannot train on video, we had to feed it to the network frame by frame.
We have chosen YOLOv5s in order to obtain balance regarding accuracy/time.
After training on 2000 frames during 10 epochs we have reached the following results:
mAP0.5 = 0.88, mAP0.5-0.95 = 0.5.
As a demo, we created a telegram bot https://t.me/TankDetection_bot that takes an image as input and sends back an image with annotated objects on it.
In the future, we plan to improve model accuracy by increasing the dataset and speeding up it by editing architecture.
Code language: Python
IDE: Colab, PyCharm
Tools: PyTorch
Space Agency Data
-
Hackathon Journey
We are entry level ML engineers and want to improve our skills in this field. Recently we have found this hackaton and this is good opportunity to practice our skills. At the same time we really want to support our country in this terrible time. That's why we decided to choose our topic - millitary vehicle detection.
We practice our skill with real life machine leraning problem and learn how to work with very small dataset.
Ukraine win!
References
Tags
#software #Ukraine #war