PSRR-MaxpoolNMS++: Fast Non-Maximum Suppression With Discretization and Pooling

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Non-maximum suppression (NMS) is critical for object detection but the standard GreedyNMS algorithm is a performance bottleneck due to its non-parallelizable nature.
• Existing parallelizable alternatives like MaxpoolNMS have limitations, restricting their use to specific stages in certain object detection architectures.

Purpose of the Study:

• To develop a generic and parallelizable NMS approach that can replace GreedyNMS in all stages of all object detectors.
• To enhance the efficiency and accuracy of parallelizable NMS methods.

Main Methods:

• Introduced a Relationship Recovery module and a Pyramid Shifted MaxpoolNMS module to improve upon MaxpoolNMS's discretization and local score calculation.
• Developed PSRR-MaxpoolNMS++ by incorporating Density-based Discretization and Adjacent Scale Pooling for more accurate suppression and efficient duplicate box identification.
• Extended PSRR-MaxpoolNMS to PSRR-MaxpoolNMS++ for enhanced performance.

Main Results:

• PSRR-MaxpoolNMS and PSRR-MaxpoolNMS++ significantly outperform MaxpoolNMS.
• PSRR-MaxpoolNMS++ achieves competitive accuracy and superior efficiency compared to the standard GreedyNMS.
• The proposed methods demonstrate effectiveness across all stages of object detection.

Conclusions:

• PSRR-MaxpoolNMS and PSRR-MaxpoolNMS++ offer viable, parallelizable alternatives to GreedyNMS.
• PSRR-MaxpoolNMS++ presents a highly efficient and accurate NMS solution suitable for all object detection pipelines.
• The developed modules enable a generic and scalable NMS replacement.