Standard Greedy Non Maximum Suppression Optimization for Efficient and High speed Inference
Seong Bin Choi, Sang-Seol Lee, Jong-Hee Park, Sung‐Joon Jang · 2021
Recent studies to improve the performance of non maximum suppression (NMS) have accuracy fluctuations when applied only to inference of models trained with standard greedy NMS. In this paper, we propose an optimization method of standard greedy NMS without degradation of accuracy. The proposed method combines NMS input and output characteristics to reduce NMS loop operations, supports high-speed inference of CNN-based object detection, and is optimized for data pipelining in hardware and CPU implementations. Experiments show that the speed of NMS in inference is improved by about ×6 compared to standard greedy NMS.