Um estudo comparativo de modelos de segmentação de objetos

Abstract

Object segmentation is the task of identifying, with pixel precision, all the objects on an image. It has several applications in the real world, such as vision systems for autonomous cars and monitoring systems. The goal of this paper is to compare three different approaches of solving that problem – Mask R-CNN, Yolact and SOLO –, all relatively recent and based on deep learning, and present some points that can be explored in future researches. By executing the models on 200 images of the OpenImages dataset, we showed that SOLO achieves the best results in terms of quality, with 42.4 of average precision (AP), about 4 points ahead of the runner-up, and Yolact, in terms of speed, with approximately half the inference time of the second place. It was also possible to observe some limitations in the area, such as the lack of annotated data on the dataset and cases that are ignored during the evaluation (more specifically, objects of classes that do not have annotations on a given image are ignored durante the AP calculations). All the code is available at https://github.com/GabrielEug2/tcc-instance-segm. Hopefully those points can be further explored in future researches, to advance the research on the area even more.