Plastic Optics

Using computer vision and satellite images to detect landfills and their environmental severity. We use the Aerial Waste Dataset provided by the Italian Agriculture Development Agency

Experiment setup

We use a two stage approach, where we first determine whether a given image is a landfill or not and also, apply segmentation to segment every object from the image and classify it according to the predefined categories.

Models

Classification models used for stage one include; Resnet18, 34, EfficientnetB0, B3

Segmentation task was based on the Deep Lab v3 Plus architecture with a Resnext50_32x4d backbone

For the second stage, we use a mult-task approach, where we both the severity and and site location.

For the multi-task, we use the same backbones are the ones for the first stage classification task

Data augmentation

We apply heavy augmentation especially the ones related to weather thanks to albumentations library

The techniques were used include;

• Random Rain

• Random Snow

• Random Fog

• Random Shadown

• Shift Scale Rotation

• RGB Shift

Training was done on Kaggle, using Pytorch and Pytorch Lightning. Weights and Biases was used for experiment tracking