TensorMask
lastmod 2026-11-10created 2019-10-263 edits
TensorMask is a deep learning model that performs instance segmentation with a dense sliding-window approach, without region proposals. It structures the output that predicts a mask at each location as a 4D tensor and introduces a tensor pyramid that varies the mask resolution by scale, achieving accuracy similar to the region-based Mask R-CNN with a single-stage architecture. A project implementation is included in Detectron2.