Performance Analysis of QUBO-translated Non-maximum Suppression for Object Detection

The industrial applicability of quantum-based systems is a focal topic in product engineering. However, critical bottlenecks for quantum-based systems include latency and technical complexity. Such challenges are evident in computer vision applications such as object detection systems. This work analyzes the possibility of applicability of a quantum computing integration in an object detection system, specifically using YOLOv7. The non-maximum suppression (NMS) algorithm was reformulated as a Quadratic Unconstrained Binary Optimization (QUBO) model for quantum applicability. The QUBO NMS model was used to filter predicted bounding boxes. A performance analysis was done to determine the applicability of a QUBO NMS by determining its mean average precision (mAP) and inference time. Experimental results show that the mAP converges at approximately 0.375 for QUBO NMS confidence thresholds from 0.1 to 0.5 at a 0.1 increment while the best average inference time at 0.5 confidence threshold is at 218 ms per image, which is a 92.35% improvement from classical NMS. The results imply that using QUBO NMS has the potential to improve an object detection system due to minimal decay in mAP compared to classical NMS when increasing confidence thresholds. However, future improvements in the penalty function may improve the mAP and inference time trade-off of QUBO-based NMS algorithms.