Transmission towers and electrical cables instance segmentation

This repository contains the code used to train and test a Mask R-CNN model for instance segmentation of transmission towers and electrical cables. The images were shot in different angles and present some obstacles that are covering the subjects. Moreover, it didn't consist of many photos, thus it was necessary to implement data augmentation techniques, like Color Jitter, rotation, cropping and a new method introduced by Ghiasi et al. in their 2020 paper "Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation". The trained model obtained a mAP50% of the bounding boxes equal to 65.8 and a mAP50% of the segmentation masks equal to 43.3. The model was trained using PyTorch and the data augmentation was implemented with Torchvision. The masks were processed with the library PyCOCOTools. In this repository there's also a notebook in which we show the results of our model. The model is in the 'checkpoints' folder.