[Paper Review] Confluence: A Robust Non-IoU Alternative to Non-Maxima Suppression in Object Detection

Andrew Shepley, Greg Falzon+1 more|arXiv (Cornell University)|Dec 1, 2020

Advanced Image and Video Retrieval Techniques80 references7 citations

TL;DR

Confluence proposes a non-IoU alternative to Non-Maxima Suppression (NMS) in object detection that uses a normalized Manhattan distance-based proximity metric to identify and suppress redundant bounding boxes. It improves average precision by up to 3.8% and average recall by up to 7.2% over state-of-the-art NMS variants on MS COCO and CrowdHuman benchmarks.

ABSTRACT

Confluence is a novel non-Intersection over Union (IoU) alternative to Non-Maxima Suppression (NMS) in bounding box post-processing in object detection. It overcomes the inherent limitations of IoU-based NMS variants to provide a more stable, consistent predictor of bounding box clustering by using a normalized Manhattan Distance inspired proximity metric to represent bounding box clustering. Unlike Greedy and Soft NMS, it does not rely solely on classification confidence scores to select optimal bounding boxes, instead selecting the box which is closest to every other box within a given cluster and removing highly confluent neighboring boxes. Confluence is experimentally validated on the MS COCO and CrowdHuman benchmarks, improving Average Precision by up to 2.3-3.8% and Average Recall by up to 5.3-7.2% when compared against de-facto standard and state of the art NMS variants. Quantitative results are supported by extensive qualitative analysis and threshold sensitivity analysis experiments support the conclusion that Confluence is more robust than NMS variants. Confluence represents a paradigm shift in bounding box processing, with potential to replace IoU in bounding box regression processes.

Motivation & Objective

• To address the limitations of IoU-based NMS in crowded or occluded scenes where high overlap causes false positive suppression.
• To overcome the reliance of NMS on classification confidence scores, which poorly correlate with localization accuracy.
• To develop a more stable and consistent method for selecting optimal bounding boxes in object detection post-processing.
• To provide a plug-and-play alternative to NMS that requires no retraining or architectural changes.

Proposed method

• Confluence uses a normalized Manhattan distance metric to measure proximity between bounding box borders, capturing geometric coherence.
• It selects the bounding box closest to all others in a cluster as the optimal representative, minimizing spatial divergence.
• Boxes are suppressed only if they are highly confluent (geometrically coherent) and redundant, not merely overlapping.
• The method does not rely on IoU or confidence scores for selection, instead prioritizing spatial alignment of box edges.
• It integrates seamlessly into existing object detectors without retraining or architectural modifications.
• Threshold sensitivity analysis confirms robustness to occlusion and varying clustering conditions.

Experimental results

Research questions

• RQ1 Can a non-IoU metric outperform IoU in suppressing redundant bounding boxes in object detection?
• RQ2 Does replacing confidence score dependence with geometric coherence improve localization accuracy and recall?
• RQ3 How does Confluence perform under high occlusion and dense object clustering compared to G-NMS and S-NMS?
• RQ4 Is Confluence more robust to hyperparameter thresholds than IoU-based NMS variants?
• RQ5 Can Confluence be deployed without retraining or model modification in real-world detection pipelines?

Key findings

• Confluence improves average precision by up to 3.8% and average recall by up to 7.2% over G-NMS and S-NMS on MS COCO and CrowdHuman.
• It achieves higher recall by preserving geometrically coherent boxes even when they are highly overlapping, unlike IoU-based methods.
• Threshold sensitivity analysis shows Confluence is more robust to hyperparameter variation than IoU-based NMS variants.
• In occluded scenarios, such as crowded scenes or partial object overlap, Confluence retains more true positives while suppressing false positives effectively.
• The method successfully avoids suboptimal box selection seen in NMS, as demonstrated in cases where IoU suppresses better-localized boxes.
• Confluence demonstrates superior performance across multiple object detectors, indicating broad compatibility and generalization.