SCNet: Training Inference Sample Consistency for Instance Segmentation

Cascaded architectures have brought significant performance improvement in object detection and instance segmentation. However, there are lingering issues regarding the disparity Intersection-over-Union (IoU) distribution of samples between training inference. This can potentially exacerbate accuracy. paper proposes an architecture referred to as Sample Consistency Network (SCNet) ensure that IoU at time is close inference time. Furthermore, SCNet incorporates feature relay utilizes global contextual information further reinforce reciprocal relationships among classifying, detecting, segmenting sub-tasks. Extensive experiments on standard COCO dataset reveal effectiveness proposed method over multiple evaluation metrics, including box AP, mask speed. In particular, while running 38\% faster, improves AP predictions by respectively 1.3 2.3 points compared strong Cascade Mask R-CNN baseline. Code available https://github.com/thangvubk/SCNet.