An encoder‐decoder framework with dynamic convolution for weakly supervised instance segmentation

In the systems of industrial robotics and autonomous vehicles, instance segmentation is widely employed. However, manually labelling an object outline time-consuming. order to reduce annotation costs, we present a weakly supervised method in this article. A deeply convolutional network first used construct multi-scale feature maps for each input image. After that, encoder-decoder framework with dynamic convolution utilised enhance model capacity efficiency, while avoiding issues anchor design, proposal selection, RoIAlign implementation. particular, Dynamic Heads are encoder create kernels, Instance decoder provide global map. With convolution, can be segmented independently, reducing interference other instances improving accuracy. Under supervision projection loss pixel point colour pairing loss, contours finally outlined. On PASCAL VOC MS COCO datasets, proposed competitive more sophisticated approaches. dataset, performance achieved 37.6% average precision ResNet-101 FPN networks. The extensively visualised results demonstrate effectiveness convolution.