Box-Supervised Instance Segmentation with Level Set Evolution

In contrast to the fully supervised methods using pixel-wise mask labels, box-supervised instance segmentation takes advantage of simple box annotations, which has recently attracted a lot research attentions. this paper, we propose novel single-shot approach, integrates classical level set model with deep neural network delicately. Specifically, our proposed method iteratively learns series sets through continuous Chan-Vese energy-based function in an end-to-end fashion. A SOLOv2 is adapted predict instance-aware map as for each instance. Both input image and its features are employed data evolve curves, where projection obtain initial boundary. By minimizing differentiable energy function, optimized within corresponding bounding annotation. The experimental results on four challenging benchmarks demonstrate leading performance approach robust various scenarios. code available at: https://github.com/LiWentomng/boxlevelset .