Back

Soft-NMS — Improving Object Detection with One Line of Code

Navaneeth BodlaBharat SinghRama ChellappaLarry S. Davis

article2017ICCV2,129 citations

Proposes a drop-in replacement for greedy non-maximum suppression that continuously decays overlapping detection scores, boosting object detector accuracy across standard benchmarks without retraining or added computational overhead.

Listen

Modern computer vision applications—such as autonomous driving, surveillance, and automated image indexing—depend heavily on accurate object detection. For decades, detection pipelines have relied on traditional non-maximum suppression to eliminate duplicate predictions around detected items. However, traditional suppression uses a strict overlap threshold that sets neighboring detection scores entirely to zero, which frequently causes real objects in crowded scenes to be missed and lowers detection accuracy.

This article proposes and evaluates "Soft-NMS," an alternative algorithm that continuously decays confidence scores as a function of bounding-box overlap rather than completely eliminating nearby detections. The goal is to retain true detections with recalibrated confidence scores while avoiding additional false positives.

To test this approach, the authors integrated linear and Gaussian score-decay formulations into leading detection frameworks, including Faster-RCNN, R-FCN, and Deformable-RFCN. They evaluated performance across standard computer vision benchmarks—specifically PASCAL VOC 2007 and MS-COCO—without retraining existing models or adding computational overhead, maintaining an identical quadratic computational complexity to traditional methods.

Empirical findings demonstrate consistent improvements across standard benchmarks. Soft-NMS improved average precision by 1.7 percentage points for Faster-RCNN and R-FCN on PASCAL VOC 2007. On the MS-COCO dataset, the method improved precision by 1.1 percentage points for Faster-RCNN, 1.3 percentage points for R-FCN, and pushed single-model Deformable-RFCN performance from 39.8% to 40.9%. The largest gains—between 3% and 6%—occurred in scenes with dense object clusters, such as herds of animals, while gains were smaller (around 0.5%) on single-stage detectors like SSD and YOLOv2 that generate fewer overlapping proposals. Across varying evaluation thresholds, Gaussian decay showed broad parameter stability around its standard setting.

These results establish that detection accuracy can be enhanced immediately without altering neural network architectures, retraining models, or increasing latency. Soft-NMS reduces missed detections in dense environments while preserving low false-positive rates, making it an efficient, low-risk upgrade for production systems that rely on multi-stage proposal detectors.

Organizations deploying object detection pipelines should replace standard non-maximum suppression with Soft-NMS, specifically using the Gaussian decay setting with its default parameter value. For single-stage detectors or domains with isolated objects, engineering teams should conduct localized pilot tests to determine if the modest gains justify deployment.

While confidence in these findings is high due to consistent performance across standard benchmarks and multiple architectures, the technique remains a greedy heuristic rather than a mathematically guaranteed global optimization. Decision-makers should note that benefits are heavily concentrated in crowded scenes with high object overlap and provide negligible impact on images with isolated, non-overlapping items.

University of Marylanddetection algorithmhttpobject detectionSoft-NMS algorithmobject detectiontraditional greedy NMS algorithm
arXiv: 1704.04503bharatsingh430/soft-nms
Cover for Soft-NMS — Improving Object Detection with One Line of Code

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. Using Deformable-RFCN, Soft-NMS improves state-of-the-art in object detection from 39.8% to 40.9% with a single model. 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 (this http URL).

Topics

detection algorithmhttpobject detectionnetwork management systems

Keywords

Soft-NMS algorithmobject detectiontraditional greedy NMS algorithmpedestrian detectionMS-COCO datasetconvolutional networks

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 Background
  • 4 Soft-NMS
  • 5 Datasets and Evaluation
  • 6 Experiments
  • 6.1 Results
  • 6.2 Sensitivity Analysis
  • 6.3 When does Soft-NMS work better?
  • 6.4 Qualitative Results
  • References

Knowls

  1. Knowl 1 — Soft-NMS Algorithm

    algorithm

    Soft Non-Maximum Suppression (Soft-NMS) is an iterative post-processing algorithm for object detection that rescores neighboring candidate bounding boxes according to a continuous decay function of their spatial overlap with the highest-scoring detection, rather than setting their scores completely to zero as in traditional greedy NMS.

    Input: Initial bounding boxes B={b1,…,bN}B = \{b_1, \dots, b_N\}, initial detection scores S={s1,…,sN}S = \{s_1, \dots, s_N\}, overlap decay function f(iou(M,bi))f(\text{iou}(M, b_i))
    Output: Final detection set DD, updated scores SS
    D←∅D \leftarrow \emptyset
    while B≠∅B \neq \emptyset do
        m←argmax⁡(S)m \leftarrow \operatorname{argmax}(S)
        M←bmM \leftarrow b_m
        D←D∪{M}D \leftarrow D \cup \{M\}
        B←B∖{M}B \leftarrow B \setminus \{M\}
        for bi∈Bb_i \in B do
            si←si⋅f(iou(M,bi))s_i \leftarrow s_i \cdot f(\text{iou}(M, b_i))
        end for
    end while
    return D,SD, S

    In each iteration, the algorithm identifies the box MM with the maximum score in the remaining candidate pool BB, moves MM to the final output set DD, and scales the confidence score sis_i of every other remaining box bi∈Bb_i \in B by f(iou(M,bi))f(\text{iou}(M, b_i)). Boxes whose decayed scores drop below a practical pruning threshold (e.g., 10−410^{-4} or 10−210^{-2}) can be discarded in each step to accelerate execution. The total computational complexity for NN candidate boxes is O(N2)\mathcal{O}(N^2), which matches greedy NMS.

  2. Knowl 2 — Gaussian Rescoring Function for Soft-NMS

    equation

    To avoid discontinuous jumps in bounding box rankings caused by hard overlap thresholds, Soft-NMS defines a smooth, continuous Gaussian penalty function for rescoring detection boxes bib_i that overlap with the currently selected maximum-scoring box MM:

    si=siexp⁡(−iou(M,bi)2σ),∀bi∉Ds_i = s_i \exp\left(-\frac{\text{iou}(M, b_i)^2}{\sigma}\right), \quad \forall b_i \notin D

    where si∈[0,1]s_i \in [0, 1] is the classification confidence score of candidate detection box bib_i, MM is the bounding box with the highest score in the current iteration, iou(M,bi)∈[0,1]\text{iou}(M, b_i) \in [0, 1] represents the Intersection-over-Union overlap between MM and bib_i, DD is the set of already confirmed detections, and σ>0\sigma > 0 is a scalar bandwidth hyperparameter (typically set to σ=0.5\sigma = 0.5).

    This continuous formulation exerts near-zero penalty when the IoU is small, gradually ramps up the decay as overlap increases, and heavily penalizes detections whose overlap with MM approaches 1, preventing sudden shifts in the relative rank of adjacent proposals.

  3. Knowl 3 — Linear Rescoring Function for Soft-NMS

    equation

    The linear rescoring function decays detection scores as a linear function of spatial overlap once a minimum overlap threshold is exceeded:

    si={si,iou(M,bi)<Ntsi(1−iou(M,bi)),iou(M,bi)≥Nts_i = \begin{cases} s_i, & \text{iou}(M, b_i) < N_t \\ s_i \left(1 - \text{iou}(M, b_i)\right), & \text{iou}(M, b_i) \ge N_t \end{cases}

    where si∈[0,1]s_i \in [0, 1] is the detection score of candidate box bib_i, MM is the bounding box with the current highest score, iou(M,bi)∈[0,1]\text{iou}(M, b_i) \in [0, 1] denotes the Intersection-over-Union overlap between MM and bib_i, and Nt∈[0,1]N_t \in [0, 1] is the overlap threshold parameter (typically set to 0.30.3). Detections with an overlap below NtN_t remain untouched, while detections with iou(M,bi)≥Nt\text{iou}(M, b_i) \ge N_t receive a penalty proportional to their overlap.

  4. Knowl 4 — Detection Performance on the MS-COCO Benchmark

    data/table

    On the MS-COCO test-dev dataset (80 object categories), replacing standard greedy NMS with Soft-NMS (using either Gaussian weighting GG with σ=0.5\sigma=0.5 or linear weighting LL with Nt=0.3N_t=0.3) yields consistent improvements across multiple object detector architectures without retraining.

    Method Training data Testing data AP 0.5:0.95 AP @ 0.5 AP small AP medium AP large Recall @ 100
    R-FCN train+val35k test-dev 31.1 52.5 14.4 34.9 43.0 43.6
    R-FCN + S-NMS G train+val35k test-dev 32.4 53.4 15.2 36.1 44.3 52.0
    R-FCN + S-NMS L train+val35k test-dev 32.2 53.4 15.1 36.0 44.1 51.0
    F-RCNN train+val35k test-dev 24.4 45.7 7.9 26.6 37.2 37.1
    F-RCNN + S-NMS G train+val35k test-dev 25.5 46.6 8.8 27.9 38.5 45.3
    F-RCNN + S-NMS L train+val35k test-dev 25.5 46.7 8.8 27.9 38.3 45.5
    D-RFCN trainval test-dev 37.4 59.6 17.8 40.6 51.4 48.3
    D-RFCN S-NMS G trainval test-dev 38.4 60.1 18.5 41.6 52.5 53.8
    D-RFCN + MST trainval test-dev 39.8 62.4 22.6 42.3 52.2 52.9
    D-RFCN + MST + S-NMS G trainval test-dev 40.9 62.8 23.3 43.6 53.3 60.4

    Soft-NMS improves COCO-style average precision (AP [0.5:0.95]) by 1.3% on R-FCN, 1.1% on Faster R-CNN (F-RCNN), 1.0% on Deformable R-FCN (D-RFCN), and 1.1% on D-RFCN with multi-scale testing (MST), pushing the single-model state of the art to 40.9%. Soft-NMS also produces substantial gains in Recall@100 (+8.4% for R-FCN and +7.5% for D-RFCN+MST).

  5. Knowl 5 — Detection Performance on the PASCAL VOC 2007 Benchmark

    data/table

    Evaluating standard off-the-shelf detectors on the PASCAL VOC 2007 test set using COCO-style average precision metrics demonstrates that Soft-NMS provides an improvement of 1.7% AP over standard greedy NMS for both Faster R-CNN and R-FCN.

    Method AP [email protected] aero bike bird boat bottle bus car cat chair cow
    Faster R-CNN + NMS 37.7 70.0 37.8 44.6 34.7 24.4 23.4 50.6 50.1 45.1 25.1 42.6
    Faster R-CNN + S-NMS G 39.4 71.2 40.2 46.6 36.7 25.9 24.9 51.9 51.6 48.0 25.3 44.5
    Faster R-CNN + S-NMS L 39.4 71.2 40.3 46.6 36.3 27.0 24.2 51.2 52.0 47.2 25.3 44.6
    R-FCN + NMS 49.8 79.4 52.8 54.4 47.1 37.6 38.1 63.4 59.4 62.0 35.3 56.0
    R-FCN + S-NMS G 51.4 80.0 53.8 56.0 48.3 39.9 39.4 64.7 61.3 64.7 36.3 57.0
    R-FCN + S-NMS L 51.5 80.0 53.2 55.8 48.9 40.0 39.6 64.6 61.5 65.0 36.3 56.5
    Method table dog horse mbike person plant sheep sofa train tv
    Faster R-CNN + NMS 36.5 40.7 46.8 39.8 38.2 17.0 37.8 36.4 43.7 38.9
    Faster R-CNN + S-NMS G 37.3 42.6 49.0 42.2 41.6 17.7 39.2 39.3 45.9 37.5
    Faster R-CNN + S-NMS L 37.2 45.1 48.3 42.3 42.3 18.0 39.4 37.1 45.0 38.7
    R-FCN + NMS 38.9 59.0 54.5 50.5 47.6 24.8 53.3 52.2 57.4 52.7
    R-FCN + S-NMS G 40.2 60.6 55.5 52.1 50.7 26.5 53.8 53.5 59.3 53.8
    R-FCN + S-NMS L 40.2 61.3 55.6 52.9 50.3 26.2 54.3 53.6 59.5 53.9

    Overall AP [0.5:0.95] increases from 37.7% to 39.4% on Faster R-CNN and from 49.8% to 51.4% (Gaussian) / 51.5% (Linear) on R-FCN, showing performance gains across nearly all individual object categories.

  6. Knowl 6 — Hyperparameter Sensitivity and Overlap Threshold Dynamics

    data/table

    Sensitivity analysis on the MS-COCO minival set comparing traditional NMS threshold NtN_t and Soft-NMS Gaussian parameter σ\sigma across different evaluation IoU thresholds Ot∈{0.5,0.6,0.7,0.8}O_t \in \{0.5, 0.6, 0.7, 0.8\} using R-FCN demonstrates that Soft-NMS provides superior performance and robustness across all evaluation criteria.

    NtN_t [email protected] [email protected] [email protected] [email protected] σ\sigma [email protected] [email protected] [email protected] [email protected]
    0.3 0.5193 0.4629 0.3823 0.2521 0.1 0.5202 0.4655 0.3846 0.2533
    0.4 0.5238 0.4680 0.3840 0.2524 0.3 0.5293 0.4765 0.3960 0.2619
    0.5 0.5227 0.4708 0.3869 0.2526 0.5 0.5274 0.4777 0.3997 0.2669
    0.6 0.5127 0.4690 0.3895 0.2527 0.7 0.5232 0.4757 0.4001 0.2695
    0.7 0.4894 0.4535 0.3860 0.2535 0.9 0.5186 0.4727 0.3992 0.2710
    0.8 0.4323 0.4048 0.3569 0.2520 1.1 0.5136 0.4691 0.3976 0.2713

    In traditional NMS, increasing NtN_t yields minor gains at high OtO_t (e.g., Ot=0.8O_t=0.8) but causes substantial drops at low OtO_t (e.g., [email protected] drops from 0.5238 at Nt=0.4N_t=0.4 to 0.4323 at Nt=0.8N_t=0.8), reducing overall multi-threshold AP. In contrast, Soft-NMS performs stably across a wide range of σ∈[0.3,0.7]\sigma \in [0.3, 0.7]; higher σ\sigma substantially improves AP at high OtO_t (reaching 0.2713 at Ot=0.8O_t=0.8) without severely degrading performance at low OtO_t.

  7. Knowl 7 — Precision-Recall and Localization Gain Mechanism of Soft-NMS

    empirical result

    Soft-NMS achieves its performance gains primarily by improving precision at higher recall levels and higher localization evaluation thresholds (OtO_t).

    In traditional greedy NMS, whenever a true object bounding box overlaps with a higher-scoring detection above the threshold NtN_t, its score is set to zero, permanently eliminating it from the ranked candidate list and imposing an upper bound on recall. Soft-NMS retains these overlapping boxes by only attenuating their confidence scores. Consequently:

    1. At lower recall levels, precision is unchanged or slightly lower because decayed boxes are ranked below unattenuated high-confidence detections.
    2. At higher recall levels and higher evaluation IoU thresholds (e.g., Ot=0.7O_t = 0.7 and Ot=0.8O_t = 0.8), Soft-NMS exhibits large precision gains over greedy NMS because true positive boxes close to other objects remain retrievable in the ranked list rather than being lost as misses.

  8. Knowl 8 — Category-Specific Efficacy of Soft-NMS

    empirical result

    The performance improvement provided by Soft-NMS varies systematically across object classes based on their real-world spatial co-occurrence and occlusion patterns:

    1. High-gain classes (3%–6% AP gain): Object categories that frequently appear in dense groups, herds, or crowded visual contexts (such as zebra, giraffe, sheep, elephant, horse, person, and motorcycle) experience the largest AP gains on MS-COCO. These categories frequently generate overlapping true positive boxes that standard NMS inadvertently suppresses.
    2. Low-gain classes (<0.5% AP gain): Object categories that rarely co-occur in close spatial proximity within the same image (such as toaster, hair drier, and sports ball) show minimal performance differences between greedy NMS and Soft-NMS.

  9. Knowl 9 — Differential Impact on Proposal-Based vs Proposal-Free Object Detectors

    empirical result

    Soft-NMS yields larger empirical gains on two-stage proposal-based object detection frameworks (such as Faster R-CNN and R-FCN) than on single-stage, proposal-free architectures (such as SSD and YOLOv2):

    1. For Faster R-CNN and R-FCN, Soft-NMS improves COCO-style AP by 1.1% to 1.7%.
    2. For single-stage detectors such as SSD and YOLOv2, linear Soft-NMS improves COCO-style AP by approximately 0.5%.

    This differential arises because two-stage proposal networks generate a high density of candidate region proposals around objects, leading to higher initial proposal recall where score decay can effectively preserve occluded true positives that hard thresholding would otherwise eliminate.

  10. Knowl 10 — Computational Efficiency and Confidence Pruning in Soft-NMS

    model/method

    Although the worst-case computational complexity of Soft-NMS is O(N2)\mathcal{O}(N^2) for NN candidate bounding boxes (identical to standard greedy NMS), its practical execution time is minimized by discarding detections whose scores drop below a small confidence threshold in each iteration.

    Empirical runtime benchmarks for Soft-NMS across 80 object classes show:

    • With a score pruning threshold of 10−410^{-4}, the per-image inference overhead is approximately 0.01 seconds using 4 CPU threads.
    • With a score pruning threshold of 10−210^{-2}, the per-image inference overhead is 0.005 seconds on a single CPU core.

    Because the score thresholding efficiently limits the active candidate set in each iteration, Soft-NMS introduces negligible computational overhead into standard object detection pipelines.

Coverage note — None was omitted; all key algorithmic formulations, mathematical scoring rules, benchmark evaluation tables, sensitivity analyses, and architectural comparisons are captured.

References

  1. 1.J. Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679–698, 1986.
  2. 2.R. T. Collins, A. J. Lipton, T. Kanade, H. Fujiyoshi, D. Duggins, Y. Tsin, D. Tolliver, N. Enomoto, O. Hasegawa, P. Burt, et al. A system for video surveillance and monitoring. 2000.
  3. 3.J. Dai, H. Qi, Y. Xiong, Y. Li, G. Zhang, H. Hu, and Y. Wei. Deformable convolutional networks. arXiv preprint arXiv:1703.06211, 2017.
  4. 4.N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on, volume 1, pages 886–893. IEEE, 2005.
  5. 5.C. Desai, D. Ramanan, and C. C. Fowlkes. Discriminative models for multi-class object layout. International journal of computer vision, 95(1):1–12, 2011.
  6. 6.P. Dollár, C. Wojek, B. Schiele, and P. Perona. Pedestrian detection: A benchmark. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 304–311. IEEE, 2009.
  7. 7.M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303–338, 2010.
  8. 8.P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627–1645, 2010.
  9. 9.A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231–1237, 2013.
  10. 10.R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014.
  11. 11.I. Haritaoglu, D. Harwood, and L. S. Davis. W/sup 4: real-time surveillance of people and their activities. IEEE Transactions on pattern analysis and machine intelligence, 22(8):809–830, 2000.
  12. 12.C. Harris and M. Stephens. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10–5244. Citeseer, 1988.
  13. 13.K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.
  14. 14.Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675–678. ACM, 2014.
  15. 15.D. Lee, G. Cha, M.-H. Yang, and S. Oh. Individualness and determinantal point processes for pedestrian detection. In European Conference on Computer Vision, pages 330–346. Springer, 2016.
  16. 16.Y. Li, K. He, J. Sun, et al. R-fcn: Object detection via region-based fully convolutional networks. In Advances in Neural Information Processing Systems, pages 379–387, 2016.
  17. 17.T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740–755. Springer, 2014.
  18. 18.W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision, pages 21–37. Springer, 2016.
  19. 19.D. G. Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91–110, 2004.
  20. 20.K. Mikolajczyk and C. Schmid. Scale & affine invariant interest point detectors. International journal of computer vision, 60(1):63–86, 2004.
  21. 21.D. Mrowca, M. Rohrbach, J. Hoffman, R. Hu, K. Saenko, and T. Darrell. Spatial semantic regularisation for large scale object detection. In Proceedings of the IEEE international conference on computer vision, pages 2003–2011, 2015.
  22. 22.J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman. Object retrieval with large vocabularies and fast spatial matching. In Computer Vision and Pattern Recognition, 2007. CVPR’07. IEEE Conference on, pages 1–8. IEEE, 2007.
  23. 23.J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 779–788, 2016.
  24. 24.S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91–99, 2015.
  25. 25.A. Rosenfeld and M. Thurston. Edge and curve detection for visual scene analysis. IEEE Transactions on computers, 100(5):562–569, 1971.
  26. 26.R. Rothe, M. Guillaumin, and L. Van Gool. Non-maximum suppression for object detection by passing messages between windows. In Asian Conference on Computer Vision, pages 290–306. Springer, 2014.
  27. 27.S. Rujikietgumjorn and R. T. Collins. Optimized pedestrian detection for multiple and occluded people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3690–3697, 2013.
  28. 28.J. Sivic, A. Zisserman, et al. Video google: A text retrieval approach to object matching in videos. In iccv, volume 2, pages 1470–1477, 2003.
  29. 29.P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple features. In Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on, volume 1, pages I–I. IEEE, 2001.
  30. 30.J. Zhang, S. Sclaroff, Z. Lin, X. Shen, B. Price, and R. Mech. Unconstrained salient object detection via proposal subset optimization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5733–5742, 2016.
  31. 31.C. L. Zitnick and P. Dollár. Edge boxes: Locating object proposals from edges. In European Conference on Computer Vision, pages 391–405. Springer, 2014.

Citation

MLA
Bodla, N., et al. “Soft-NMS -- Improving Object Detection With One Line of Code”. arXiv, 2017, http://arxiv.org/abs/1704.04503v2.
APA
Bodla, N., Singh, B., Chellappa, R., & Davis, L. S. (2017). Soft-NMS -- Improving Object Detection With One Line of Code. arXiv. http://arxiv.org/abs/1704.04503v2
Chicago
Bodla, N., B. Singh, R. Chellappa, and L. S. Davis. 2017. “Soft-NMS -- Improving Object Detection With One Line of Code”. arXiv. http://arxiv.org/abs/1704.04503v2.
Harvard
Bodla, N. et al. (2017) “Soft-NMS -- Improving Object Detection With One Line of Code”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1704.04503v2.
Vancouver
1. Bodla N, Singh B, Chellappa R, Davis LS (2017) Soft-NMS -- Improving Object Detection With One Line of Code. arXiv

BibTeX

@article{bodla2017soft,
  title = {Soft-NMS -- Improving Object Detection With One Line of Code},
  author = {Bodla, Navaneeth and Singh, Bharat and Chellappa, Rama and Davis, Larry S.},
  year = {2017},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1704.04503v2},
  eprint = {1704.04503}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: IEEE