# Non-maximum suppression

> Source: https://aiwiki.ai/wiki/non_maximum_suppression
> Updated: 2026-11-18
> Fact-checked: 2026-11-18
> Categories: Algorithms, Computer Vision, Deep Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Non-maximum suppression." aiwiki.ai, 18 Nov 2026. https://aiwiki.ai/wiki/non_maximum_suppression
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Non-maximum suppression** (**NMS**) is a post-processing algorithm that removes redundant, overlapping detections so that each object in an image is reported once. In [object detection](https://aiwiki.ai/wiki/object_detection) it sorts candidate [bounding boxes](https://aiwiki.ai/wiki/bounding_box) by confidence score, keeps the highest-scoring box, discards every remaining box whose overlap with it exceeds a threshold, and repeats on what is left [1]. The overlap is usually measured with [intersection over union](https://aiwiki.ai/wiki/iou) (IoU) [5][21]. NMS was for many years a standard final stage of detectors from the deformable part model to the [R-CNN](https://aiwiki.ai/wiki/r_cnn) family, [YOLO](https://aiwiki.ai/wiki/yolo) and SSD [2][4][5][6][8].

The idea is older than deep learning. Bodla and colleagues, writing in 2017, described NMS as part of detection algorithms in [computer vision](https://aiwiki.ai/wiki/computer_vision) "for almost 50 years", first in edge detection, where it thins edges by removing spurious responses, and later in feature point detection, face detection and object detection [1]. Its fixed overlap threshold forces a trade-off between missing objects that genuinely overlap and keeping duplicate boxes, which led to a series of variants (Soft-NMS, DIoU-NMS, adaptive NMS, parallel matrix forms, learned NMS networks) and to detectors, such as [DETR](https://aiwiki.ai/wiki/detr) (2020) and YOLOv10 (2024), that are trained to produce one prediction per object and need no NMS at inference [1][2][15][18].

## Why detectors produce duplicates

Most detectors score a large set of candidate windows, anchors or proposals more or less independently. Two windows that cover almost the same image region see almost the same pixels, so both receive a high score, and a single object ends up triggering several detections of varying confidence [2]. Bounding-box regression adds to the clutter: in two-stage detectors, several proposals are often regressed onto the same region of interest [1]. Hosang, Benenson and Schiele summarized the resulting pipeline as three steps: propose a search space of windows, score and refine each window, then merge windows that may belong to the same object, the last step being what is called non-maximum suppression [2].

Evaluation is the other reason NMS matters. Benchmarks such as [PASCAL VOC](https://aiwiki.ai/wiki/pascal_voc) and [COCO](https://aiwiki.ai/wiki/coco_dataset) count only one detection per ground-truth object as a true positive, so extra boxes on an already detected object are scored as false positives [1][2]. The DETR authors described the problem as collapsing "near-duplicate predictions", and noted that the performance of conventional detectors is significantly influenced by that post-processing step [15].

<figure class="embedvideo">
<iframe src="https://www.youtube-nocookie.com/embed/VAo84c1hQX8" title="C4W3L07 Nonmax Suppression"></iframe>
<figcaption>"C4W3L07 Nonmax Suppression", video from DeepLearningAI, published November 7, 2017. <a href="https://www.youtube.com/watch?v=VAo84c1hQX8">Watch on YouTube</a>.</figcaption>
</figure>

## History

According to Bodla et al., NMS was first used in edge detection, citing Rosenfeld and Thurston's 1971 paper on edge and curve detection, with later edge-detection uses including Canny's 1986 detector [1]. In feature point detectors such as the Harris corner detector it acts as local thresholding that yields unique feature points [1]. In the Viola and Jones face detector, overlapping detections were partitioned into disjoint subsets by an overlap criterion and the final box was the average of each subset's coordinates [1]. Hosang et al. trace the general practice of collapsing highly overlapping detections into one to at least 1994, citing a face detection paper by Burel and Carel [2].

Bodla et al. credit Dalal and Triggs' 2005 human detector with showing that a greedy procedure (select the highest-scoring box, suppress its neighbours above an overlap threshold) worked better than the face-detection approach, after which greedy NMS became the de facto algorithm for object detection [1]. The deformable part model of Felzenszwalb, Girshick, McAllester and Ramanan sorted detections by score and greedily kept the best ones, skipping any detection whose box was at least 50% covered by a previously selected box [4]. R-CNN applied greedy NMS to each class independently, rejecting a region if its IoU with a higher-scoring selected region exceeded a learned threshold [5].

Deep single-stage and two-stage detectors kept the step. The table below lists settings reported in some of the best-known papers.

| Detector | Year | Where NMS is applied | Reported setting |
|---|---|---|---|
| [R-CNN](https://aiwiki.ai/wiki/r_cnn) | 2014 | Scored region proposals, per class | Greedy, IoU threshold learned [5] |
| [Faster R-CNN](https://aiwiki.ai/wiki/faster_r_cnn) | 2015 | Region proposal network outputs | IoU 0.7, leaving about 2,000 proposals per image [7] |
| [YOLO](https://aiwiki.ai/wiki/yolo) (v1) | 2015 | Final detections | Not critical, but adds 2-3% mAP [6] |
| SSD | 2016 | Final detections, per class | Confidence 0.01, IoU 0.45, top 200 per image [8] |
| R-FCN and Faster R-CNN (as tested by Bodla et al.) | 2017 | Final detections, per class | IoU 0.3 default [1] |

In the SSD paper, NMS on SSD300 with the 20 PASCAL VOC classes cost about 1.7 ms per image, close to the 2.4 ms spent on all of the network's newly added layers [8]. Faster R-CNN's authors reported that applying NMS to the proposals did not harm final detection accuracy while substantially reducing the number of proposals [7].

## Algorithm

Greedy NMS takes a list of boxes $B = \{b_1, \dots, b_N\}$ with scores $S = \{s_1, \dots, s_N\}$ and an overlap threshold $N_t$ [1]:

1. Select the box $M$ with the highest remaining score, remove it from $B$ and add it to the output set $D$.
2. Remove from $B$ every box $b_i$ with $\mathrm{IoU}(M, b_i) \ge N_t$.
3. Repeat until $B$ is empty [1].

The IoU of two boxes $A$ and $B$ is the area of their intersection divided by the area of their union, $|A \cap B| / |A \cup B|$ [9]. Bodla et al. wrote the pruning step as a rescoring rule in which a neighbour's score is kept when its overlap with $M$ is below the threshold and set to zero otherwise:

$$s_i = \begin{cases} s_i, & \mathrm{IoU}(M, b_i) < N_t \\ 0, & \mathrm{IoU}(M, b_i) \ge N_t \end{cases}$$

Each iteration compares $M$ with up to $N$ remaining boxes, so the procedure is $O(N^2)$ in the number of boxes [1]. In practice the number of boxes is reduced first with a confidence threshold, and the run time then depends on how many boxes survive and on both thresholds [1][17].

### Practical details

- **Per-class or class-agnostic.** Detectors usually run NMS separately for each class, so a box of one class never suppresses a box of another [1][5]. The torchvision library for [PyTorch](https://aiwiki.ai/wiki/pytorch) exposes this as `batched_nms`, which takes a category index for each box and does not apply NMS between elements of different categories [22]. The Ultralytics YOLO library has an `agnostic_nms` option that suppresses overlapping boxes across classes instead [20].
- **Threshold boundary.** Implementations differ on whether a box exactly at the threshold is removed. The ONNX `NonMaxSuppression` operator suppresses boxes whose IoU with a selected box is strictly greater than `iou_threshold`, so a box at exactly the threshold is kept [23]; torchvision's `nms` likewise discards boxes with IoU greater than the threshold [21].
- **Ties.** torchvision documents that when several boxes have exactly the same score, the box selected is not guaranteed to be the same on CPU and GPU [21].
- **Output caps.** Pipelines often keep only the top-scoring detections after NMS, for example the top 200 per image in SSD [8]. As of November 2026 the Ultralytics prediction defaults are a confidence threshold of 0.25, an NMS IoU threshold of 0.5 and at most 300 detections per image [20].

## The threshold trade-off

A single fixed threshold cannot suit every scene. Hosang et al. argued that greedy NMS works only if suppression is wide enough to remove all high-scoring duplicates of one object yet narrow enough never to remove a high-scoring detection of the next closest object; in crowded scenes with heavy occlusion these two conditions conflict, so the algorithm "is doomed to sacrifice precision or recall independent of its parameter" [2]. In their earlier paper they described the failure modes directly: a threshold that is too strict leaves high-scoring [false positives](https://aiwiki.ai/wiki/false_positive), while one that is too loose merges true positives and lowers recall [3].

Bodla et al. made the same point in terms of the COCO metric, which averages [average precision](https://aiwiki.ai/wiki/average_precision) over overlap thresholds from 0.5 to 0.95. A low NMS threshold such as 0.3 can delete a box that would have matched an object at a strict evaluation overlap, while a high one such as 0.7 adds false positives at loose evaluation overlaps [1]. Liu, Huang and Wang illustrated the problem on crowded pedestrian images: with Faster R-CNN detections, an NMS threshold of 0.5 removed true positives, while 0.7 let more false positives through [10].

## Variants

### Soft-NMS

Soft-NMS, published by Navaneeth Bodla, Bharat Singh, Rama Chellappa and Larry S. Davis of the University of Maryland at ICCV 2017, replaces deletion with score decay [1]. Instead of setting a neighbour's score to zero, it lowers the score as a function of the neighbour's overlap with $M$, so a box covering a second, overlapping object stays in the ranked list with reduced confidence [1]. The paper proposed a linear rule,

$$s_i = \begin{cases} s_i, & \mathrm{IoU}(M, b_i) < N_t \\ s_i \,(1 - \mathrm{IoU}(M, b_i)), & \mathrm{IoU}(M, b_i) \ge N_t \end{cases}$$

and, because the linear rule is discontinuous at $N_t$, a Gaussian rule applied to all remaining boxes at every iteration,

$$s_i = s_i \, e^{-\mathrm{IoU}(M, b_i)^2 / \sigma}$$

where $\sigma$ controls how quickly the penalty grows with overlap [1]. In the paper's experiments $N_t$ was 0.3 for the linear rule and $\sigma$ was 0.5 for the Gaussian rule [1]. Greedy NMS is the special case with a binary weighting function, and Soft-NMS keeps the same $O(N^2)$ complexity [1].

The authors reported gains in COCO-style mAP of 1.7 points on PASCAL VOC 2007 for both R-FCN and Faster R-CNN, and of 1.3 points (R-FCN) and 1.1 points (Faster R-CNN) on MS-COCO, with no retraining. With Deformable R-FCN and multi-scale testing, Soft-NMS raised COCO test-dev AP from 39.8% to 40.9% [1]. The authors described the method as a single-line modification of the greedy algorithm, which gave the paper its subtitle [1].

### DIoU-NMS

Zheng et al. proposed Distance-IoU (DIoU) in a paper accepted to AAAI 2020, mainly as a bounding-box regression loss, and also suggested it as a suppression criterion [9]. DIoU subtracts a penalty $\rho^2(b, b^{gt}) / c^2$ from IoU, where $\rho$ is the Euclidean distance between the two box centres and $c$ is the diagonal length of the smallest box enclosing both [9]. DIoU-NMS suppresses a box when $\mathrm{IoU} - \rho^2/c^2$ reaches the threshold, on the reasoning that two boxes with distant centres probably locate different objects and should not be removed even when they overlap [9].

### Adaptive NMS

Adaptive NMS, by Songtao Liu, Di Huang and Yunhong Wang of Beihang University (CVPR 2019), targets pedestrian detection in crowds [10]. A small subnetwork predicts an object density $d_M$ for each detection, defined during training as the box's maximum IoU with any other ground-truth box. The suppression threshold around a selected box becomes $N_M = \max(N_t, d_M)$: in sparse regions this is ordinary NMS, while in crowded regions the threshold rises so that neighbouring boxes, which probably belong to other people, survive [10]. The authors reported state-of-the-art results on the CityPersons and CrowdHuman pedestrian benchmarks [10].

### Parallel NMS: Fast NMS and Matrix NMS

Greedy NMS is sequential, because whether a box survives depends on which higher-scoring boxes survived before it. The YOLACT instance segmentation paper (Bolya et al., ICCV 2019) introduced Fast NMS to remove that dependency [11]. It computes a pairwise IoU matrix for the top $n$ detections of each class, zeroes the lower triangle and diagonal, takes the column-wise maximum, and keeps the detections whose maximum overlap with any higher-scoring detection is below the threshold [11]. Already-suppressed boxes are thereby allowed to suppress others, which removes slightly too many boxes, but every decision can be made in parallel with standard GPU matrix operations. The authors measured Fast NMS as 11.8 ms faster than a Cython implementation of traditional NMS at a cost of 0.1 mAP, and 15.0 ms faster than the CUDA implementation in the Mask R-CNN benchmark suite at a cost of 0.3 mAP [11].

Matrix NMS, introduced with SOLOv2 (Wang et al., NeurIPS 2020), is a parallel, score-decaying method for instance masks that its authors say was motivated by Soft-NMS, where computing IoU between masks is more expensive than between boxes [12]. For each prediction it computes a decay factor from the penalties imposed by all higher-scoring predictions, each compensated by how strongly that higher-scoring prediction is itself overlapped by predictions above it (an approximation of its chance of being suppressed), and the whole computation is done in one shot from an $N \times N$ IoU matrix [12]. The SOLOv2 authors reported that it processed 500 masks in under 1 ms in a simple Python implementation and scored 0.4% AP higher than Fast NMS [12].

### Learned NMS

Several groups tried to replace the hand-written rule with a neural network. Hosang, Benenson and Schiele first proposed the "Tyrolean network", a convnet that performs NMS by rescoring a given set of detections, and tested it on synthetic data and crowded pedestrian scenes [3], then, at CVPR 2017, a "GossipNet" that rescores detections using only their boxes and scores, with the aim of producing exactly one high-scoring detection per object and making the final greedy step unnecessary [2]. Averaged across the 80 COCO classes, one multi-class GossipNet reached 24.3% mAP (over IoU thresholds 0.5 to 0.95), compared with 23.5% for a greedy NMS tuned per class on the test set [2].

Relation Networks for Object Detection (Hu et al., Microsoft Research Asia and Peking University, CVPR 2018) added an attention-based object relation module that lets detections interact through their appearance and geometry [13]. Used for duplicate removal, a lightweight relation network replaced NMS and, according to the authors, surpassed Soft-NMS; they described the resulting system as the first fully end-to-end object detector [13].

### Box fusion for ensembles

NMS only selects among existing boxes. Weighted boxes fusion (Solovyev, Wang and Gabruseva, published in Image and Vision Computing in 2021) instead combines the predictions of several detection models into averaged boxes weighted by their confidence scores, a method aimed at [ensembles](https://aiwiki.ai/wiki/ensemble_learning) where real-time inference is not required [14]. The authors reported top results with it in the Open Images and COCO detection tracks [14].

| Method | Year | Change from greedy NMS | Source |
|---|---|---|---|
| Soft-NMS | 2017 | Decays neighbour scores (linear or Gaussian) instead of deleting them | [1] |
| GossipNet | 2017 | Neural network rescores detections from boxes and scores | [2] |
| Relation network duplicate removal | 2018 | Attention module learns duplicate removal inside the detector | [13] |
| Fast NMS | 2019 | Parallel matrix form; suppressed boxes may suppress others | [11] |
| Adaptive NMS | 2019 | Threshold raised per box using predicted crowd density | [10] |
| DIoU-NMS | 2020 | Suppression criterion also penalizes centre distance | [9] |
| Matrix NMS | 2020 | Parallel score decay for instance masks, motivated by Soft-NMS | [12] |
| Weighted boxes fusion | 2021 | Averages boxes from several models instead of selecting one | [14] |

## Implementations and deployment cost

NMS is available as a built-in operator in common deep learning tooling. torchvision provides `torchvision.ops.nms`, which takes boxes in `(x1, y1, x2, y2)` format, scores and an IoU threshold and returns the indices of kept boxes sorted by decreasing score [21]. [ONNX](https://aiwiki.ai/wiki/onnx) defines a `NonMaxSuppression` operator, available since opset version 10, with inputs for boxes, per-class scores, a maximum number of output boxes per class, and IoU and score thresholds [23]. NVIDIA's [TensorRT](https://aiwiki.ai/wiki/tensorrt) shipped an Efficient NMS plugin, intended mainly for EfficientDet but usable with SSD or Faster R-CNN, that can fuse box decoding into the NMS operation; its ONNX variant is deprecated since TensorRT 9.0, superseded by a built-in NMS layer [24].

Because the time NMS takes depends on how many boxes survive the confidence filter, it is hard to benchmark. In their 2023 RT-DETR paper (CVPR 2024), Zhao et al. measured the TensorRT EfficientNMS kernel with YOLOv8 on an NVIDIA T4 GPU and found that its execution time grows as the confidence threshold falls or the IoU threshold rises; YOLOv8 reached its best AP with a confidence threshold of 0.001 and an IoU threshold of 0.7, where NMS time was relatively high [17]. They noted that YOLO papers typically report model speed without NMS time, built an end-to-end speed benchmark that includes it, and found that anchor-based YOLO detectors produced about three times more boxes, and so needed more NMS time, than anchor-free ones [17].

## NMS-free detectors

The alternative to better NMS is a detector that does not produce duplicates in the first place. Hosang et al. argued that a detector can output one detection per object only if it processes candidate detections jointly, so that it can tell when several of them describe the same object [2].

### DETR and set prediction

[DETR](https://aiwiki.ai/wiki/detr) (DEtection TRansformer), from Facebook AI (Carion et al., ECCV 2020), treats detection as direct set prediction [15]. A [transformer](https://aiwiki.ai/wiki/transformers) decoder outputs a fixed-size set of predictions, and during training the Hungarian algorithm finds a one-to-one bipartite matching between predictions and ground-truth objects on which the loss is computed, so the model is trained to produce unique predictions and does not need NMS by design [15]. The authors checked this by running standard NMS on the output of each decoder layer. NMS helped the first layer, which cannot compare its predictions with one another, but its benefit shrank with depth, and at the last layers it slightly reduced AP by removing true positives, since [self-attention](https://aiwiki.ai/wiki/self_attention) between predictions lets the model inhibit duplicates itself [15].

Later work extended the approach. DeFCN (Wang et al., CVPR 2021) showed that a fully convolutional detector could drop NMS with a prediction-aware one-to-one label assignment and a 3D max filtering module, reporting that the one-to-one assignment alone obtained performance comparable with NMS and that the full end-to-end framework was competitive with many state-of-the-art NMS-based detectors on COCO and CrowdHuman [16]. RT-DETR, from researchers at Baidu and Peking University, described itself as the first real-time end-to-end detector, motivated by the observation that NMS hurts both the speed and the accuracy of YOLO models [17].

### YOLOv10 and later YOLO models

YOLOv10 (Wang et al., Tsinghua University, NeurIPS 2024) brought NMS-free inference to the [YOLO](https://aiwiki.ai/wiki/yolo) family [18]. Its authors observed that YOLO models are trained with one-to-many label assignment, in which one ground-truth object has several positive predictions, and therefore depend on NMS at inference, which slows inference and makes performance sensitive to NMS hyperparameters [18]. Their "consistent dual assignments" add a second, one-to-one head with the same structure: both heads are trained together so the backbone still gets the richer one-to-many supervision, and at inference the one-to-many head is discarded and the one-to-one head makes the predictions without NMS [18]. A shared matching metric $m = s \cdot p^{\alpha} \cdot \mathrm{IoU}(\hat{b}, b)^{\beta}$, where $p$ is the classification score and $s$ a spatial prior, keeps the two heads' assignments consistent [18].

Ultralytics' YOLO26, which the company says was released in January 2026, uses a similar dual-head design [19]. As of November 2026 its documentation describes the one-to-many head, which requires NMS, as the default because it is slightly more accurate, with the NMS-free one-to-one head available by setting `nms=False` and capped at 300 detections [19]. When NMS-free inference is used, Ultralytics' class-agnostic option only removes the same detection appearing with several class labels and performs no IoU-threshold suppression between distinct boxes [20].

## See also

- [IoU](https://aiwiki.ai/wiki/iou)
- [Object detection](https://aiwiki.ai/wiki/object_detection)
- [Bounding box](https://aiwiki.ai/wiki/bounding_box)
- [DETR](https://aiwiki.ai/wiki/detr)
- [YOLO](https://aiwiki.ai/wiki/yolo)
- [Post-processing](https://aiwiki.ai/wiki/post-processing)
- [Average precision](https://aiwiki.ai/wiki/average_precision)

## References

1. Navaneeth Bodla, Bharat Singh, Rama Chellappa, Larry S. Davis. "Soft-NMS -- Improving Object Detection With One Line of Code". ICCV 2017 (arXiv:1704.04503), 2017. https://arxiv.org/abs/1704.04503
2. Jan Hosang, Rodrigo Benenson, Bernt Schiele. "Learning Non-Maximum Suppression". Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. https://openaccess.thecvf.com/content_cvpr_2017/html/Hosang_Learning_Non-Maximum_Suppression_CVPR_2017_paper.html
3. Jan Hosang, Rodrigo Benenson, Bernt Schiele. "A convnet for non-maximum suppression". arXiv:1511.06437, 2015. https://arxiv.org/abs/1511.06437
4. Pedro F. Felzenszwalb, Ross B. Girshick, David McAllester, Deva Ramanan. "Object Detection with Discriminatively Trained Part-Based Models". IEEE Transactions on Pattern Analysis and Machine Intelligence 32(9):1627-1645, September 2010. https://doi.org/10.1109/TPAMI.2009.167
5. Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik. "Rich feature hierarchies for accurate object detection and semantic segmentation". CVPR 2014 (arXiv:1311.2524), 2013. https://arxiv.org/abs/1311.2524
6. Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi. "You Only Look Once: Unified, Real-Time Object Detection". arXiv:1506.02640, 2015. https://arxiv.org/abs/1506.02640
7. Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun. "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks". arXiv:1506.01497, 2015. https://arxiv.org/abs/1506.01497
8. Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg. "SSD: Single Shot MultiBox Detector". ECCV 2016 (arXiv:1512.02325), 2015. https://arxiv.org/abs/1512.02325
9. Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, Dongwei Ren. "Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression". AAAI 2020 (arXiv:1911.08287), 2019. https://arxiv.org/abs/1911.08287
10. Songtao Liu, Di Huang, Yunhong Wang. "Adaptive NMS: Refining Pedestrian Detection in a Crowd". CVPR 2019 (arXiv:1904.03629), 2019. https://arxiv.org/abs/1904.03629
11. Daniel Bolya, Chong Zhou, Fanyi Xiao, Yong Jae Lee. "YOLACT: Real-time Instance Segmentation". ICCV 2019 (arXiv:1904.02689), 2019. https://arxiv.org/abs/1904.02689
12. Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, Chunhua Shen. "SOLOv2: Dynamic and Fast Instance Segmentation". NeurIPS 2020 (arXiv:2003.10152), 2020. https://arxiv.org/abs/2003.10152
13. Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, Yichen Wei. "Relation Networks for Object Detection". Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. https://arxiv.org/abs/1711.11575
14. Roman Solovyev, Weimin Wang, Tatiana Gabruseva. "Weighted boxes fusion: Ensembling boxes from different object detection models". Image and Vision Computing, article 104117, 2021 (arXiv:1910.13302). https://arxiv.org/abs/1910.13302
15. Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko. "End-to-End Object Detection with Transformers". Computer Vision - ECCV 2020, Lecture Notes in Computer Science (arXiv:2005.12872), 2020. https://arxiv.org/abs/2005.12872
16. Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, Nanning Zheng. "End-to-End Object Detection with Fully Convolutional Network". CVPR 2021 (arXiv:2012.03544), 2020. https://arxiv.org/abs/2012.03544
17. Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, Jie Chen. "DETRs Beat YOLOs on Real-time Object Detection". Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16965-16974. https://openaccess.thecvf.com/content/CVPR2024/html/Zhao_DETRs_Beat_YOLOs_on_Real-time_Object_Detection_CVPR_2024_paper.html
18. Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, Guiguang Ding. "YOLOv10: Real-Time End-to-End Object Detection". NeurIPS 2024 (arXiv:2405.14458), 2024. https://arxiv.org/abs/2405.14458
19. Ultralytics. "Ultralytics YOLO26". Ultralytics Docs, accessed November 2026. https://docs.ultralytics.com/models/yolo26/
20. Ultralytics. "Model Prediction with Ultralytics YOLO". Ultralytics Docs, accessed November 2026. https://docs.ultralytics.com/modes/predict/
21. Torch Contributors. "torchvision.ops.nms". Torchvision documentation, accessed November 2026. https://docs.pytorch.org/vision/stable/generated/torchvision.ops.nms.html
22. Torch Contributors. "torchvision.ops.batched_nms". Torchvision documentation, accessed November 2026. https://docs.pytorch.org/vision/stable/generated/torchvision.ops.batched_nms.html
23. ONNX. "NonMaxSuppression". ONNX Operators documentation, accessed November 2026. https://onnx.ai/onnx/operators/onnx__NonMaxSuppression.html
24. NVIDIA. "Efficient NMS Plugin". NVIDIA TensorRT GitHub repository (release/10.0), accessed November 2026. https://github.com/NVIDIA/TensorRT/tree/release/10.0/plugin/efficientNMSPlugin

