Computer Vision News - May 2020

2 Summary Deep Learning Library 0 by Marica Muffoletto Applications of imgaug Library for Augmentation Hi all! Here we are again with a new article on deep learning and a tool that can help us develop our models. While in the last two months we focused on the problem of explainability, we will now approach a topic that is indeed closely linked to it: uncertainty. To further dig into the concept of deep learning as a black box, we will see how we can use Test-Time Augmentation (TTA) to explore uncertainty and hence to improve the understanding of the predictions. We will do that through a specific library called imgaug. imgaug is an excellent python library available for augmentation purposes and obtainable through pip or conda installation. It can be adopted for many deep learning applications, ranging from the most obvious one, increasing the variance in the training sample, to most sophisticated ones such as Test-Time Augmentation. The library allows to dynamically and randomly apply transformations of any type to not only images, but also keypoints, bounding boxes or heatmaps. It includes a wide range of both shape and intensity augmenters. As we will show in the second section of the article, it also offers the functionality to apply the same transformation to both an image and its mask, making it a significant tool in applications such as segmentation or object detection algorithms. Application 1: test-time augmentation Let’s first start by importing all the libraries we need and an already trained model from pytorch and set it to eval mode. At this step, we can load any segmentation/ object detection model that we would like to test. Here we used a multiclass segmentation model trained on a subset of the COCO dataset available here . The architecture offered by pytorch is called Deeplabv3 and uses a ResNet-101 backbone. The accuracy of the pretrained model results in a mean IoU of 67.4.