Reversible Recursive Instance-Level Object Segmentation

摘要: In this work, we propose a novel Reversible Recursive Instance-level Object Segmentation (R2-IOS) framework to address the challenging instance-level object segmentation task. R2-IOS consists of reversible proposal refinement sub-network that predicts bounding box offsets for refining locations, and an generates foreground mask dominant instance in each proposal. By being recursive, iteratively optimizes two subnetworks during joint training, which refined proposals improved predictions are alternately fed into other progressively increase network capabilities. reversible, adaptively determines optimal number iterations required both training testing. Furthermore, handle multiple overlapped instances within proposal, instance-aware denoising autoencoder is introduced distinguish from distracting instances. Extensive experiments on PASCAL VOC 2012 benchmark well demonstrate superiority over state-of-the-art methods. particular, APr 20 classes at 0:5 IoU achieves 66:7%, significantly outperforms results 58:7% by PFN [17] 46:3% [22].