In modern object detection pipelines, Non-Maximum Suppression (NMS) acts as the final arbiter. Its job is to take a dense cluster of candidate bounding boxes and filter out redundant duplicates, leaving exactly one box per object.
Traditionally, NMS relies entirely on Intersection over Union (IoU). If a candidate box overlaps with a higher-confidence box beyond a certain area threshold, standard NMS assumes it is a redundant detection and deletes it.
While this works well for isolated objects, it fails catastrophically in crowded or highly occluded scenes. Consider two vehicles bumper-to-bumper in a traffic jam, or tightly packed items on a retail shelf. In these scenarios, distinct objects naturally exhibit high spatial overlap. Standard IoU-based NMS looks only at the overlapping area, mistakes these distinct objects for duplicates, and erroneously suppresses valid detections. This creates a fundamental trade-off between filtering out true false positives and preserving valid, closely packed detections.
To resolve this ambiguity, a powerful paradigm shift has emerged: Distance-Aware and Structure-Aware Non-Maximum Suppression. By incorporating spatial proximity, geometric dimensions, and internal structural landmarks into the suppression metric, models can differentiate between redundant boxes targeting the same object and highly overlapping boxes targeting distinct objects.
The Core Intuition: Center-Point Proximity
How do we mathematically distinguish between two separate objects that happen to overlap, and two redundant candidate boxes drawn around the exact same object?
The key insight is that even when two bounding boxes share a large portion of their area, their centroids behave differently depending on what they contain. If two boxes are predicting the same object, their center points should be nearly coincident. If they are predicting two distinct, adjacent objects, their centers will remain spatially offset.
Redundant Detections (Same Object) Distinct Overlapping Objects
+-----------------+ +---------+-------+
| +-----------+ | | | |
| | * | | | * | * |
| +-----------+ | | | |
+-----------------+ +---------+-------+
Centers are highly coincident. Centers are distinctly offset.
By measuring the Euclidean distance between bounding box centers, we can penalize or adjust the suppression criteria. In their seminal paper on Distance-IoU (DIoU), Zheng et al. introduced DIoU-NMS within Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression. Instead of evaluating pure IoU, the suppression metric subtracts a normalized distance penalty based on the central points of the two boxes.
If the centers are far apart, the distance penalty reduces the effective overlap score below the suppression threshold. As a result, the neighboring box is preserved. This simple geometric modification has been widely adopted to prevent the accidental removal of valid bounding boxes in dense object clusters, significantly boosting recall in domains like vehicle tracking (Zhao et al. in Exploration of Vehicle Target Detection Method Based on Lightweight YOLOv5 Fusion Background Modeling), maritime ship detection (Xiao et al. in A Tiny Model for Fast and Precise Ship Detection via Feature Channel Pruning), and retail commodity tracking (Xie et al. in Detection of Commodities Based on Multi-Feature Fusion and Attention Screening by Entropy Function Guidance).
Refining Geometry: Accounting for Scale and Aspect Ratio
Center distance alone, however, can still struggle when objects are occluded along the same line of sight—for instance, a small object directly in front of a larger object, where their centroids align perfectly but their physical boundaries differ.
To handle these complex occlusions, researchers have expanded the geometric criteria to include box dimensions:
- Scale and Aspect Alignment: Li et al., in RE-YOLOv5: Enhancing Occluded Road Object Detection via Visual Receptive Field Improvements, utilize EIOU-NMS. This method builds on center distance by adding explicit penalties for differences in bounding box width and height. If two boxes share a center but have vastly different dimensions, they are recognized as separate entities and protected from suppression.
- Dimensional Ratios: In multi-view environments like roadside LiDAR, Lan et al. proposed Center Distance IoU (CDIoU) within 3D Point Cloud Stitching for Object Detection with Wide FoV Using Roadside LiDAR. CDIoU balances center proximity with dimensional ratios to accurately deduplicate overlapping fields of view without sacrificing true positives.
- Localization Precision: Similarly, Teng et al. developed a position-sensitive mechanism in MSR-RCNN: A Multi-Class Crop Pest Detection Network Based on a Multi-Scale Super-Resolution Feature Enhancement Module to prevent the model from selecting poorly centered, high-overlap boxes in dense agricultural pest clusters.
- Layout-Driven Filtering: For highly structured, dense layouts like document tables, the authors of Scene Table Structure Recognition with Segmentation and Key Point Collaboration designed Centroid Filtering-based NMS (CF-NMS) to resolve adjacent, overlapping table cell candidates where standard area metrics frequently fail.
Structural Proxies: Keypoints and Pixel-Level Overlaps
When bounding boxes are too coarse to capture the true orientation or shape of an object, distance-aware NMS can look inside the box to evaluate structural landmarks or pixel distributions.
Object Keypoint Similarity (OKS)
For articulated objects like humans, a bounding box is a poor proxy for spatial occupancy. Two people dancing or hugging might share the exact same bounding box coordinates, yet their physical bodies remain distinct.
To solve this, Papandreou et al. introduced an NMS strategy based on Object Keypoint Similarity (OKS) in Towards Accurate Multi-person Pose Estimation in the Wild. Instead of measuring box boundaries, this approach calculates the spatial distribution and visibility of specific anatomical landmarks (e.g., shoulders, knees, eyes). If the underlying keypoints show distinct spatial configurations, the system recognizes them as two separate individuals, allowing for precise pose deduplication in highly crowded scenes.
Alternative Overlap Metrics
For highly irregular, non-convex, or fragmented shapes—such as collapsed buildings in post-earthquake drone imagery—standard IoU calculations break down because the rectangular boxes contain mostly empty background space.
To address this, Ding et al. adapted NMS to use a pixel-wise overlap metric called the Intersected Proportion of Objects (IPO) in A Precision Efficient Method for Collapsed Building Detection in Post-Earthquake UAV Images Based on the Improved NMS Algorithm and Faster R-CNN. By evaluating intersection at the structural pixel level rather than the box level, they achieved cleaner and more accurate localization. In a similar vein, Wu et al. utilized an Intersection Over Area (IOA) metric in A Coarse-to-Fine Network for Ship Detection in Optical Remote Sensing Images to better preserve valid detections in dense, parallel object clusters like ships docked side-by-side.
Summary
Standard NMS operates in a single dimension: area overlap. Distance-aware and structure-aware NMS expand this decision boundary into a multi-dimensional space. By combining area overlap with center-point distances, aspect ratios, and structural keypoints, these methods allow modern object detectors to cleanly separate redundant false positives from tightly packed, valid detections.
Key Foundational Work
A Coarse-to-Fine Network for Ship Detection in Optical Remote Sensing Images (2020) — contributed 2 concept elements to this direction.
Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression (2020) — contributed 2 concept elements to this direction.
Towards Accurate Multi-person Pose Estimation in the Wild (2017) — contributed 1 concept elements to this direction.