⭐ Weakly-supervised Rip Current Segmentation

Box-supervised Segmentation | CondInst | Detectron2 | Semi Segmentation | Object Segmentation | Computer Vision | Deep Learning

  • Achieved a 30% performance boost over fully supervised models by designing a novel dual-phase architecture.
  • Designed a box-supervised instance segmentation algorithm leveraging the CondInst as base model.
  • Implemented contrastive learning and model ensemble on augmented data for high-precision localization.


Since our paper is currently under review on a prestigeous journal, we can't reveal the actual full architecture of our framework. The above image reflects a shallow-level structure of the workflow. Our major goal is solving the tiresome work of the mask annotation, especially of the amorphous and occuled rip currents phenomena.