Ensemble Learning to Perform Instance Segmentation over Synthetic Data

Abstract

Recently, deep neural networks have led the progress of instance segmentation. There are several models of neural networks, such as the Mask R-CNN network, that perform this task with good results. However, it is possible to improve the results of the Mask R-CNN network if several models of it are combined, and their results are merged using Ensemble Learning. We propose the algorithm “Simple Instance Segmentation Ensemble” that is capable of assembling the results of two Mask R-CNN networks to produce better results. In our experiments, we train several Mask R-CNN networks with synthetic images of machinery objects. In addition, these Mask R-CNN networks have different backbones and different sizes of kernels for the Gaussian Blur filter applied to the synthetic machinery images used during training. We tested the performance of these networks by predicting real images of machinery. Besides, we propose the SISE algorithm to assemble the predictions of two previously trained Mask R-CNN networks, and we obtained better results than those of the individual Mask R-CNN networks. In particular, our best result is an ensemble that has one Mask R-CNN trained with synthetic images smoothed by Gaussian Blur filter with a kernel size of 7 $\times$ 7, and another network with a kernel size of 3 $\times$ 3. Both networks have as backbone a ResNeXt 101 with FPN (Feature Pyramid Network). This ensemble has a bounding box mAP of 89.42% and a segmentation mAP of 88.34% in the real machinery test images.