Vision Network

This project consists in training a CNN,based on the Gatenet or Dronet architecture, to detect drone racing gates and get gate size and location. For more information, consult the Nano Drone Racing repository.

Description

A streamlined CNN, based on the Gatenet architecture, has been adapted to minimize computational demand. This network is successfully deployed on a GAP8 processor, achieving a processing rate of 16Hz. The CNN provides data on gate size and location, which serves as input for the positioning algorithm

This repository can train 2 different CNN architetures GateNet and Dronet. As well as an experimental active vision network. Moreover, we can finetune a network by lock in place certain layers and augments training data.

Getting Started

Dependencies

• Conda or Pip

Installing

1. Clone the repository:

   git clone https://github.com/fed12345/visionnet
   cd visionnet

2. Set Up a Virtual Environment (recommended)

It’s a best practice to use a virtual environment to manage dependencies. To create a virtual environment, run the following command with conda installed:

conda create --name visionnet
conda activate visionnet

3. Intall Dependencies

With the environment active, install all necessary packages by running:

pip install -r requirements.txt

4. Training data

The datasets used of training are found in Gatenet

Executing Progam

Train network

To choose the architecture and train the network:

python3 train.py -c config/gatenet.json

Data Augmentation

To finetune the model:

python3 finetune.py