DataTerminology.comNEW
Also known as: NMS
A post-processing method that filters overlapping detections, keeping the highest-scoring boxes and discarding redundant ones.
DESCRIPTION
Non-maximum suppression (NMS) consolidates overlapping detections in object detection outputs. It ranks boxes by confidence and removes those with high overlap (e.g., IoU) with a selected box. NMS reduces duplicate predictions and improves precision. Variants like soft-NMS decay scores rather than hard removal. Proper IoU and score thresholds affect recall and precision trade-offs. NMS is standard in YOLO, Faster R-CNN, and other detectors. Efficient implementations are critical for real-time pipelines.
ALTERNATIVE DEFINITIONS
PyTorch torchvision
Performs non-maximum suppression (NMS) on the boxes according to their intersection-over-union (IoU).
Highlights IoU-based filtering in practical API.
OpenCV
Performs non-maximum suppression given scores and boxes.
Library-level description of NMS over detection outputs.
Wikipedia
Non-maximum suppression (NMS) is a class of algorithms used in computer vision to select one entity among overlapping entities.
General CV framing explaining selection among overlaps.
RELATED TERMS
Object Detection Intersection over Union Bounding Box Confidence Score YOLO Faster R-CNN Anchor Boxes Post-processing
COMPETING TERMS
Soft-NMS Weighted Box Fusion
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer Software Developer
EXTERNAL SOURCES
- Wikipedia - Non-maximum suppression
- Ultralytics YOLO - Explanation of NMS in YOLO
- TensorFlow - TensorFlow NMS API reference
- Bodla et al. (Soft-NMS) - Soft-NMS paper introducing score decay
TOP QUESTIONS
What problem does Non-Maximum Suppression solve?
It removes duplicate overlapping detections, improving precision and cleaner outputs in object detection.
Category: Relevancy
How does IoU relate to Non-Maximum Suppression?
IoU quantifies overlap between boxes; NMS uses IoU thresholds to discard redundant detections.
Category: Requirements
What is Soft-NMS compared to Non-Maximum Suppression?
Soft-NMS decays scores of overlapping boxes instead of removing them, often improving recall.
Category: Alternatives
What are common mistakes configuring Non-Maximum Suppression?
Using too high IoU thresholds can suppress true objects; too low thresholds can leave duplicates.
Category: Mistakes
Is Non-Maximum Suppression differentiable?
Standard NMS isn’t; differentiable approximations exist but are less common in production.
Category: Topic
How does Non-Maximum Suppression affect latency?
Efficient NMS implementations are crucial for real-time detection; naive NMS can be a bottleneck.
Category: Risks
Where is Non-Maximum Suppression applied in a detection pipeline?
It runs as post-processing on model outputs before final predictions are returned.
Category: Management
Can Non-Maximum Suppression handle multi-class detections?
Yes, it is commonly applied per class to avoid cross-class suppression issues.
Category: Requirements
What alternatives exist to improve Non-Maximum Suppression results?
Weighted box fusion combines boxes to improve localization; Soft-NMS improves recall.
Category: Alternatives