Confluence: A Robust Non-IoU Alternative to Non-Maxima Suppression in Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection models generate bounding boxes that often overlap, requiring post-processing to eliminate redundant detections.
• Traditional Non-Maxima Suppression (NMS) methods, primarily based on Intersection over Union (IoU), suffer from limitations in stability and consistency.
• Existing NMS variants like Greedy and Soft NMS rely heavily on classification scores, potentially leading to suboptimal bounding box selection.

Purpose of the Study:

• To introduce Confluence, a novel alternative to IoU-based NMS for bounding box post-processing in object detection.
• To address the limitations of existing NMS methods by proposing a more stable and consistent bounding box clustering approach.
• To evaluate the effectiveness of Confluence against established NMS variants on benchmark datasets.

Main Methods:

• Confluence utilizes a normalized Manhattan Distance-inspired proximity metric for bounding box clustering, moving beyond IoU.
• It identifies the optimal bounding box within a cluster as the one closest to all other boxes in that cluster.
• The method removes highly confluent neighboring boxes based on this proximity metric, not solely on confidence scores.

Main Results:

• Confluence demonstrated improved Average Precision (AP) by 0.2-2.7% on MS COCO and 1-3.8% on CrowdHuman.
• Average Recall (AR) saw significant gains, ranging from 1.3-9.3% on MS COCO and 2.4-7.3% on CrowdHuman, compared to Greedy and Soft-NMS.
• Extensive qualitative and threshold sensitivity analyses confirmed Confluence's robustness over NMS variants.

Conclusions:

• Confluence offers a more stable and consistent approach to bounding box clustering compared to IoU-based NMS.
• The method achieves superior performance in both Average Precision and Average Recall on challenging benchmarks.
• Confluence represents a significant advancement in bounding box processing and has the potential to replace IoU in regression tasks.