PointINS: Point-Based Instance Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Segmentation

Background:

• Instance segmentation with Point-of-Interest (PoI) features faces challenges in differentiating multiple instances within a single PoI.
• Learning high-dimensional mask features for each instance using standard convolution is computationally intensive.

Purpose of the Study:

• To propose an efficient and effective method for mask representation in instance segmentation.
• To address the computational burden of learning instance-specific mask features.

Main Methods:

• Introduced instance-aware convolution, decomposing mask representation into instance-aware weights and instance-agnostic features.
• Developed PointINS, a practical instance segmentation approach built on dense one-stage detectors.
• Utilized RetinaNet and FCOS frameworks for evaluation.

Main Results:

• PointINS with a ResNet101 backbone achieved 38.3 mask mean average precision (mAP) on the COCO dataset.
• Significantly outperformed existing point-based instance segmentation methods.
• Demonstrated comparable performance to region-based Mask R-CNN with faster inference.

Conclusions:

• Instance-aware convolution offers an efficient approach to mask representation learning in instance segmentation.
• PointINS provides a simple yet powerful framework for high-performance instance segmentation.
• The proposed method achieves state-of-the-art results for point-based approaches and competitive results against region-based methods.