[Paper Review] Improving Object Detection With One Line of Code.

Navaneeth Bodla, Bharat Singh+2 more|arXiv (Cornell University)|Apr 14, 2017

Advanced Neural Network Applications20 references95 citations

TL;DR

This paper proposes Soft-NMS, a simple yet effective modification to non-maximum suppression (NMS) that replaces hard suppression with score decay based on IoU overlap, improving mAP by 1.1–1.7% on PASCAL VOC 2007 and MS-COCO without additional training or hyperparameters. It maintains the same computational complexity as standard NMS while eliminating false positives from strict IoU-based suppression.

ABSTRACT

Non-maximum suppression is an integral part of the object detection pipeline. First, it sorts all detection boxes on the basis of their scores. The detection box M with the maximum score is selected and all other detection boxes with a significant overlap (using a pre-defined threshold) with M are suppressed. This process is recursively applied on the remaining boxes. As per the design of the algorithm, if an object lies within the predefined overlap threshold, it leads to a miss. To this end, we propose Soft-NMS, an algorithm which decays the detection scores of all other objects as a continuous function of their overlap with M. Hence, no object is eliminated in this process. Soft-NMS obtains consistent improvements for the coco-style mAP metric on standard datasets like PASCAL VOC 2007 (1.7\% for both R-FCN and Faster-RCNN) and MS-COCO (1.3\% for R-FCN and 1.1\% for Faster-RCNN) by just changing the NMS algorithm without any additional hyper-parameters. Further, the computational complexity of Soft-NMS is the same as traditional NMS and hence it can be efficiently implemented. Since Soft-NMS does not require any extra training and is simple to implement, it can be easily integrated into any object detection pipeline. Code for Soft-NMS is publicly available on GitHub \url{http://bit.ly/2nJLNMu}.

Motivation & Objective

• To address the limitation of traditional NMS, which discards overlapping detections even when they are valid, leading to missed detections.
• To improve mean average precision (mAP) in object detection without requiring additional training or hyperparameter tuning.
• To design a method that is simple, efficient, and compatible with existing object detection pipelines.
• To replace the hard suppression in NMS with a continuous, differentiable score decay mechanism based on intersection over union (IoU).

Proposed method

• Instead of removing all boxes with IoU above a threshold, Soft-NMS reduces their detection scores based on a continuous function of their overlap with the highest-scoring box.
• The score decay function is applied recursively to remaining boxes after the highest-scoring box is selected.
• The decay function is defined as a Gaussian or linear function of IoU, where higher overlap leads to greater score reduction.
• The algorithm maintains the same O(n log n) time complexity as standard NMS, enabling efficient deployment.
• Soft-NMS is fully backward compatible and can be integrated into any object detector with minimal code change—just one line of code.
• The method does not require retraining or additional hyperparameters, making it plug-and-play for existing models.

Experimental results

Research questions

• RQ1 Can a continuous score decay mechanism improve mAP in object detection compared to hard suppression in standard NMS?
• RQ2 Does Soft-NMS maintain computational efficiency while reducing false positives caused by aggressive IoU thresholding?
• RQ3 Can Soft-NMS be applied universally across different detectors like Faster R-CNN and R-FCN without retraining?
• RQ4 What is the impact of Soft-NMS on standard benchmarks such as PASCAL VOC 2007 and MS-COCO?

Key findings

• Soft-NMS improves mAP by 1.7% on PASCAL VOC 2007 for both Faster R-CNN and R-FCN.
• On MS-COCO, Soft-NMS boosts mAP by 1.3% for R-FCN and 1.1% for Faster R-CNN.
• The performance gain is consistent across multiple detectors, demonstrating broad applicability.
• The method achieves these gains without any additional training or hyperparameter tuning.
• Soft-NMS maintains the same O(n log n) computational complexity as standard NMS, ensuring efficiency.
• The code is publicly available, enabling immediate integration into existing detection pipelines.