Geração automática de mapas de dificuldade para datasets

Abstract

Difficulty maps are structures that store estimated difficulty levels for a change detection algorithm to classify each pixel in a video's frames correctly. Such maps are essential to obtain an objective measure, considering the "difficulty level" of the video. Generating a video difficulty map requires using ground truths. Creating a ground truth is considered a laborious process as it involves assigning, often manually, labels to all pixels of all video frames. In this research, a method is presented, which consists of a trained neural network capable of estimating the difficulty level of a video without using a ground truth. We expected that researchers use the results of this work to generate new videos datasets to evaluate change detection algorithms.