MambaInst: Lightweight State Space Model for Real-Time Instance Segmentation

In this paper, we propose a lightweight and efficient state-space model-based instance segmentation network named MambaInst, which extracts deep semantic features through a LightSSM Block consisting of gating mechanisms and residual connectivity to model long-distance spatial dependencies with linear computational complexity. We design a novel downsampling method called FRDown to efficiently capture contextual information, thereby improving the network’s local information perception. With its excellent model architecture and simple training method, MambaInst-B achieves 40.8% in Mask mAP on a single 4090 GPU with an inference time of 2.28 ms on the COCO-seg. Our proposal demonstrates first proof of SSM’s effectiveness in real-time instance segmentation, setting a new performance benchmark for Mamba-based techniques in this particular application.