BoundNet: Pixel Level Boundary Marking and Tracking of Instance Video Objects

Object segmentation is the inevitable result of object recognition and semantic segmentation when moving from coarse to fine inference. Despite many efforts in segmenting instances, the quality of masks remains unsatisfactory. Because of the low pixel-to-edge ratio and the restricted spatial resolution of the feature map, the boundaries of the predicted version masks are typically inaccurate. Defining the object boundaries at the detailed pixel level is essential for instance segmentation. This paper aims a simple, yet conceptually effective, mask-based RCNN framework to generate accurate segmentation masks to overcome these challenges. Additionally, it can extract the foreground objects from the background and run at high inference speeds to meet the growing demand for accurate, real-time computer vision applications. This study introduces a new database called RSBFRS that includes unconstrained images and videos. Thus instance segmentation using Mask RCNN was implemented to mark/extract precise boundaries of moving person in a video and the maximum prediction score value of 0.9983 obtained. This efficient technique can serve as a basic support for future studies on instance-level detection and recognition.