hustvl/SparseInst
A real-time instance segmentation framework based on sparse instance activation maps, achieving 37.9 AP at 40 FPS.

Velocity · 7d
+0.4
★ / day
Trend
→steady
star history
SparseInst introduces Instance Activation Maps (IAM) to adaptively highlight informative object regions for recognition. It is a fully convolutional instance segmentation framework that eliminates the need for non-maximum suppression (NMS) or sorting, making it efficient and easy to deploy. The method achieves competitive accuracy-speed trade-offs on standard benchmarks like MS COCO.