Focal Phi Loss with ACU-Net

This repository holds the source code for Focal Phi Loss, a novel loss function for highly imbalanced datasets applied to Power line detection with Auxiliary Classifier U-Net (ACU-Net), by Rabeea Jaffari, Manzoor Ahmed Hashmani and Constantino Carlos Reyes-Aldasoro.

Prerequisites

Python >= 3.6
PyTorch == 1.7, tested on CUDA 10.1. The models were trained and evaluated on PyTorch 1.7. Torch Vision==0.8.1
Torchaudio==0.7.0 MMSegmentation MMCCV ==1.1.4
Other dependencies described in requirements.txt

Install

The code in this repo is built using the mmsegmentation framework. For more information on the mmsegmentation framework see:
https://github.com/open-mmlab/mmsegmentation
https://mmsegmentation.readthedocs.io/en/latest/

Datasets

The two benchmark Power line (PL) datasets used in this research are:

1. Mendeley PL dataset available at: https://data.mendeley.com/datasets/twxp8xccsw/8
2. Power line dataset of urban scenes (PLDU) available at: https://drive.google.com/drive/folders/1XjoWvHm2I8Y4RV_i9gEd93ZP-KryjJlm

The train/val splits of these datasets used in the experiments can be found at: dataset_files

Models

1. Auxiliary Classifier U-Net (ACU-Net)
2. Vanilla U-NET

Train and Test Models on PL datasets

To train and test the ACU-Net model on:
Mendeley dataset acunet_mendeley
PLDU dataset acunet_pldu

To train and test the Vanilla U-Net model on:
Mendeley dataset vanilla_unet_mendeley
PLDU dataset vanilla_unet_pldu

License

This project is released under the Apache 2.0 license.