SPRNet: Single-Pixel Reconstruction for One-Stage Instance Segmentation.

Area of Science:

• Computer Vision
• Image Segmentation

Background:

• Object instance segmentation is crucial for pixel-level image understanding but is computationally intensive.
• Existing two-stage methods, while effective, suffer from low inference speeds, limiting practical applications.

Purpose of the Study:

• To propose an efficient one-stage framework for object instance segmentation.
• To enhance the speed and performance of existing one-stage detectors for segmentation tasks.

Main Methods:

• Introducing a single-pixel reconstruction (SPR) branch to off-the-shelf one-stage detectors.
• The SPR branch reconstructs pixel-level masks directly from convolution feature maps.
• Utilizing a ResNet-50 backbone for the proposed SPRNet framework.

Main Results:

• SPRNet achieves comparable mask Average Precision (AP) to Mask R-CNN.
• SPRNet demonstrates a higher inference speed compared to two-stage methods.
• SPRNet shows all-round improvements in box AP across all scales compared to RetinaNet.

Conclusions:

• SPRNet offers an efficient and effective solution for object instance segmentation.
• The proposed single-pixel reconstruction approach enhances the performance of one-stage detectors.
• SPRNet balances segmentation accuracy with significantly improved inference speed.