InstanceNet: object instance segmentation using DNN —— CSI6900 Project

One-stage object detectors like SSD [1] and YOLO [2] are able to speed up existing two-stage detectors like Faster R-CNN [3] by removing the object proposal stage and making up for the lost performance in other ways. Nonetheless, the same approach is not easily transferable to instance segmentation task. Current one-stage instance segmentation methods can be simply classified into segmentation-based methods which segment first then do clustering, and proposal-based methods which detect first then predict masks for each instance proposal. Proposal-based methods always enjoy a better mAP; by contrast, segmentation-based methods are generally faster when inferencing. In this work, we first propose a one-stage segmentation-based instance segmentation solution, in which a pull loss and a push loss are used for differentiating instances. We then propose two post-processing methods, which provide a trade-off between accuracy and speed.