CO²Det: Anchor-free based Convex-hull Generation Network for Oriented Object Detection

Abstract

Nowadays, arbitrary oriented object detection has achieved considerable progress in the field of remote sensing image interpretation. However, there remains challenges of label assignment and loss discontinuity. To put the axe in the helve, we propose an anchor-free Convex-hull Oriented Object Detector (CO ² Det). In CO ² Det, the convex-hull feature expression(CFE) module and Kullback-Leibler Divergence(KLD) loss are proposed, based on the 135° long edge prediction box definition. In the CFE, a convex-hull set is constructed for each object so strengthen the connection of non-axis aligned feature from background pixels or adjacent objects. To maintain the loss continuity and scale invariance, a general rotation detection loss KLD is devised that transform the convex-hull parameters into a 2-D Gaussian distribution. KLD is applied to release mutual constraints between parameters so that adjust the direction optimization strategy adaptively. The DOTA and HRSC2016 public datasets are chosen to train our model and achieved the state-of-the-art performance.