lab-cv: A Jupyter Notebook repository from naivomah3

Summary

This repository is holding experiments related to my M.Tech thesis. The project is aiming to address the road extraction problem in an end-to-end fashion. Our dataset is a set of aerial images taken from UAV(drones) from local areas within the NIT Rourkela campus. Throughout the experiments, we are benchmarking different state-of-the-art models and taking advantage of the techniques being used to tackle our problem. The main objective of this project is to build an effective CNN model, being able to distinguish roads from occlusion and background and able to generalize to later extension as well as to build our own dataset.

Please note: the development is undergoing and details will gradually be provided below.

Install required libraries

There are a few important dependencies which need to be installed before making any change in pre-processing or expermenting the models. Note that the code has been tested using python==3.6.7.

pip install -r requirements.txt

Few parameters(environment variables) have to be set according to the need.

In train.sh and predict.sh

export PRE_TRAINED=True   # set to False if not loading the pre-trained weight 
export WEIGHTS_IN_PATH=path/to/the/weights.h5    # full path of the weight

Create new conda environment

conda create --name ENV_NAME python=3.6.7
source activate ENV_NAME

Dataset directory structure

/input contains images/frames and their respective masks within each set(training/validation/testing). Within mask, the ground truth set would be a Tensor of dimension (height x width) and each pixel's value maps the spatial distribution of the respective label.

/input
├── testing # Testing data 
│   ├── images
│   └── mask
├── training   # Training data 
│   ├── images
│   └── mask
└── validation # Validation data
    ├── images
    └── mask

/output contains predicted masks referred from input/prediction along with their respective frames within each respective subfolders and named the same way to avoid confusion. masks and prediction sets would be a (3D) tensor of dimension (height x width x 3) colored with respect to the spatial distribution of each class/label. As an example, please find here

/output   
└── prediction # Prediction data 
    ├── images
    ├── mask
    └── prediction

/models is place to store and load the model weights.

Usage

For training and prediction, refer to train.sh and predict.sh respectively to define necessary environment variables.

# How to train? 
sh train.sh
# How to predict?
sh predict.sh

Timeline