Part-aware Prototype Network for Few-shot Semantic Segmentation 

      European Conference on Computer Vision (ECCV 2020)

      [paper] [code] [video] [slides] 

• Part-based prototype representation reserves more detailed information.

• Unlabeled data helps to capture object appearance variations.

• Add additional semantic branch for learning a better visual representation.

• Our method outperforms the SOTA with a sizable margin of 3.32% and 3.30% for 1-shot and 5-shot setting.

Auto-segmentation of Soft X-Ray Images reveals beta cell architectural variances during Glucose and Ex-4 stimulation

     [code] [Slides]

• Propose a novel spatio-temporal segmentation cell model of Pancreatic β-cells.

• Develop a strategy with biological analysis to present the organelles interactive behaviors under drugs stimulations in a systematic way.

• Generate a correlative model from multi biological modalities.