Higher-Order Recurrent Network With Space-Time Attention For Video Early Action Recognition

Cell and nuclei instance segmentation is crucial for computer-assisted interpretation of Cervical Cytology. However, due to the overlapping of cells and miscellaneous objects, it is not trivial to segment them jointly and bind the them accurately. in this work, we propose Bi-Polar Mask that utilizes two polars to represent the contours of cell and nucleus. The polar representation in nature overcomes the challenges of object overlapping and cell folding. We also design a shared center at the nuclei, which allows for straightforward binding the pair of contours. A large dataset with thousands of cell and nucleus annotations was constructed for evaluation. Experimental results demonstrate the superior performance of the proposed method.