Image-Level Supervised Instance Segmentation with Inter-Instance Knowledge

Image-level supervised methods are an effective way to reduce the cost of dense annotation for instance segmentation. Some effective methods have been developed in recent years. However, most existing methods treat each image individually. As we know, different instances of the same category usually share some common appearances. Motivated by this intuition, we consider inter-instance knowledge constraints to improve imagelevel supervised instance segmentation results. In this paper, an image-level supervised instance segmentation method with interinstance knowledge is developed based on BEyond Semantic segmentation To InstancE Segmentation (BESTIE) framework, named BESTIE-inter. Inter-instance knowledge is first collected and established as a codebook based on some simple samples. The collected knowledge is then used as a supervision to guide the segmentation of other instances. Experimental results on PASCAL VOC 2012 demonstrate that, compared with the baseline, BESTIE-inter has obtain a performance gain of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.3 \% \text{mAP} 50$</tex> and 2.2% mAP75, outperforming most existing state-of-the-art methods.