Improvement of Bounding Box and Instance Segmentation Accuracy Using ResNet-152 FPN with Modulated Deformable ConvNets v2 Backbone-based Mask Scoring R-CNN

Abstract

A challenging task is to make sure that the deep learning network learns prediction accuracy by itself. Intersection-over-Union (IoU) amidst ground truth and instance mask determines mask quality. There is no relationship between classification score and mask quality. The mission is to investigate this problem and learn the predicted instance mask’s accuracy. The proposed network regresses the MaskIoU by comparing the predicted mask and the respective instance feature. The mask scoring strategy determines the disorder among mask score and mask quality, then adjusts the parameters accordingly. Adaptation ability to the object’s geometric variations decides deformable convolutional network’s performance. Using increased modeling power and stronger training, focusing ability on pertinent image regions is improved by a reformulated Deformable ConvNets. The introduction of modulation technique, which broadens the deformation modeling scope, and the integration of deformable convolution comprehensively within the network enhance the modeling power. The features which resemble region-based convolutional neural network (R-CNN) feature’s classification capability and its object focus are learned by the network with the help of feature mimicking scheme of DCNv2. Feature mimicking scheme of DCNv2 guides the network training to efficiently control this enhanced modeling capability. The backbone of the proposed Mask Scoring R-CNN network is designed with ResNet-152 FPN and DCNv2 network. The proposed Mask Scoring R-CNN network with DCNv2 network is also tested with other backbones ResNet-50 and ResNet-101. Instance segmentation and object detection on COCO benchmark and Cityscapes dataset are achieved with top accuracy and improved performance using the proposed network.