基于Anchor-free的图像全景分割算法研究

IMAGE PANOPTIC SEGMENTATION BASED ON ANCHOR-FREE

对于实例分割子任务,大多数框架都采用基于anchor的策略,如著名的Mask R-CNN。但anchor依赖于人为预先设置,且由于anchor和RPN的存在,网络计算量往往很大。设计一个自定义的U-Net骨干网络和一个立足于极坐标系对实例分割掩膜边缘进行数学建模的anchor-free的实例分割网络,将二者集成到全景分割架构中去。该全景分割网络摆脱人工预设的anchor,实现轻量级,U-Net骨架网相比FCN网干网在精度相差不多的情况下参数量下降35%,更适合低开销实时应用场景。

For instance segmentation subtasks,most frameworks use anchor-based strategies,such as the well-known Mask R-CNN.But their anchors rely on artificial pre-settings,and because of the existence of anchors and RPN,the amount of network computations is often large.A custom U-Net backbone network and an anchor-free instance segmentation network which was based on polar coordinate system to mathematically model the edge of the instance segmentation mask was designed,and the two were integrated into the panoptic segmentation architecture.The panoptic segmentation network got rid of the artificial preset anchor and achieved lightweight.Compared with FCN,the custom U-Net backbone network had similar accuracy while the amount of parameters was reduced by 35%,which was more suitable for low-cost real-time application scenarios.