ARMA Networks for Image Segmentation

Authors

Arpit Aggarwal Shishira Maiya Shantam Bajpai

Introduction to the Project

ARMA stands for Auto-regressive Moving Average, a concept that was recently introduced by a research group at UMD. The aim of adding interconnections between output neurons is to increase the net receptive field which inturn helps in learning more information in the image. This is very useful for image segmentation and object detection tasks. In this project, different CNN Architectures like ARMA Deeplab-v3, Deeplab-v3, ARMA Deeplab-v3+ and Deeplab-v3+ were used for the task of image segmentation on Cityscapes dataset. The input to the CNN networks was a (768x768x3) image and the number of classes were equal to 19. The CNN architectures were implemented in PyTorch and the loss function was Cross Entropy(CE) Loss. The hyperparameters to be tuned were: Number of epochs(e), Learning Rate(lr), weight decay(wd) and batch size(bs).

Data

The dataset used was Cityscapes dataset for the task of image segmentation(number of classes=19). The dataset can be downloaded from here: https://www.cityscapes-dataset.com/downloads/

Software Required

To run the jupyter notebooks, use Python 3. Standard libraries like Numpy and PyTorch are used.

Credits

The following links were helpful for this project:

1. https://github.com/umd-huang-lab/ARMA-Networks
2. https://github.com/meetshah1995/pytorch-semseg/
3. https://github.com/bodokaiser/piwise/