CoChat
Sign Up
Cited by 543Advanced Neural Network ApplicationsAdvanced Image and Video Retrieval TechniquesVisual Attention and Saliency Detection

Weighted boxes fusion: Ensembling boxes from different object detection models

Open access
Solovyev, Roman, Wang, Weimin, Gabruseva, Tatiana·2021·Image and Vision Computing
DOI 10.1016/j.imavis.2021.104117
Open this paper in CoChat — free accountOpen PDF

AI snapshot

Written from the paper’s full text.

TL;DR

Weighted Boxes Fusion (WBF) ensembles object detection boxes by fusing all predictions using confidence-weighted coordinates and rescaling confidence by cluster size and model count. It outperformed NMS, soft-NMS, and NMW on COCO and Open Images ensembles, reaching 56.1 mAP on COCO validation and 56.4 mAP on COCO test.

Key findings

  • On a two-model COCO ensemble (EfficientDetB6 + EfficientDetB7), WBF achieved mAP(0.5..0.95) = 0.5344, mAP@0.5 = 0.7244, and mAP@0.75 = 0.5824, compared with NMS 0.5269/0.7156/0.5737, soft-NMS 0.5239/0.7121/0.5677, and NMW 0.5285/0.7171/0.5743.
  • For test-time augmentation with EfficientDetB7, WBF reached mAP(0.5..0.95) = 0.5262, versus NMS 0.5233, soft-NMS 0.5210, and NMW 0.5250.
  • On a many-model COCO ensemble, WBF gave 56.1 mAP on validation and 56.4 mAP on test, described as the top third result on the MS COCO leaderboard as of 07/08/2020.
  • On Open Images, WBF reached 0.5665 mAP@0.5 for a RetinaNet ensemble (vs NMS 0.5442 and soft-NMS 0.5428) and 0.5982 mAP@0.5 for an ensemble of fairly different models (vs NMW 0.5667, NMS 0.5642, soft-NMS 0.5616), helping achieve 7th out of 558 in the challenge.

Methods

  • WBF collects all predicted boxes, sorts them by confidence, clusters boxes with IoU > THR (THR = 0.55 was close to optimal in experiments), fuses coordinates as confidence-weighted sums, sets fused confidence as the average confidence, then rescales confidence by the number of boxes in the cluster divided by the number of models.
  • Evaluation used Open Images V5 (about 16M boxes, 600 classes, 1.9M images; training 1,743,042 images, validation 41,620, test 99,999; mAP@0.5) and MS COCO (over 200,000 images, 80 categories; train2017 118k, val2017 5k, test 20k; mAP over IoU 0.5 to 0.95).
  • Models included RetinaNet, EfficientDet, DetectoRS, YOLOv5, Faster R-CNN, Mask R-CNN, and Cascade R-CNN; WBF was compared against NMS, soft-NMS, and NMW, with grid search for optimal parameters.

Limitations

  • The paper states WBF is slower than NMS and soft-NMS, on average around three times slower than standard NMS.
  • The paper reports that replacing NMS at the output of an individual model with WBF degraded performance: RetinaNet ResNet152 on Open Images gave raw boxes 0.1718 mAP, NMS default 0.4902, NMS optimal 0.4906, and WBF optimal 0.4532; the authors attribute this to excessive low-scored wrong predictions, noting WBF works better for ensembles of fairly accurate models.
  • Evident from the text: WBF performance depends on tuned parameters such as IoU threshold and model weights, which were optimized by grid search in the reported experiments.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.imavis.2021.104117

Cites

33 of 33 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 27 cited sources checked · 0 supported · 0 flagged

  • Findingown data
    § 1. Introduction · p. 1

    WBF significantly improves the quality of the combined predicted rectangles.

    “This method significantly improves the quality of the combined predicted rectangles.”
    • Also stated in: § 6.2. An ensemble of two different models · p. 4 · boosted
      “Even for the ensemble of only two models, the WBF method outperformed significantly other methods.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 6.4.2 Ensemble of RetinaNet models for Open Images Dataset · p. 5

    The WBF method demonstrated the best results among the compared methods for combining RetinaNet predictions on the Open Images dataset.

    “the WBF method demonstrated the best results.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 6.4.2 Ensemble of RetinaNet models for Open Images Dataset · p. 6

    The paper's results are the third top entry in the COCO Detection Challenge as of 07/08/20.

    “The third top entry represents the results described in this paper.”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 6.4.3 Ensemble of fairly different models for Open Images Dataset · p. 6

    The WBF method outperformed by far the other algorithms (NMS, soft-NMS, NMW) for combining predictions of fairly different models on the Open Images dataset.

    “The WBF method outperformed by far the other algorithms (see Table 5)”
    Quote in the paperSubstantial claim
  • Findingown data
    § 6.4.2 Ensemble of RetinaNet models for Open Images Dataset · p. 5

    Combining predictions from several RetinaNet detectors with different backbones yields better performance than individual models on the Open Images dataset.

    “Combining predictions from several detectors with different backbones yields better performance”
    Quote in the paperSubstantial claim
  • May combine two claims. 13
  • Classified as probably not a substantial claim. 15
  • The quote only partly states this claim. 3

  • May combine two claims.

    Findingown data
    § 6.4.1 Ensemble of models for COCO Dataset · p. 5

    The WBF method for combining predictions of detection models gave 56.1 mAP for the validation data set and 56.4 mAP for the test set on MS COCO.

    “The WBF method for combining predictions of detection models gave 56.1 mAP for the validation data set and 56.4 mAP for the test set.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 8. Conclusion · p. 7

    WBF achieved the top performance for the COCO Detection Challenge, with 56.1 mAP for validation and 56.4 mAP for test-dev.

    “the top performance for the COCO Detection Challenge (( 56.1 mAP for validation data set and 56.4 mAP for the test-dev set).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data
    § 3. Weighted Boxes Fusion · p. 2

    In our experiments, THR = 0.55 was close to an optimal threshold for WBF.

    “Note: in our experiments, THR = 0 . 55 was close to an optimal threshold.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data
    § 6.2. An ensemble of two different models · p. 4

    Even for the ensemble of only two models, the WBF method outperformed significantly other methods (NMS, soft-NMS, NMW) on MS COCO validation set.

    “Even for the ensemble of only two models, the WBF method outperformed significantly other methods.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given
    § 7. Discussion · p. 7

    WBF is on average around three times slower than standard NMS.

    “the WBF algorithm, on average, is around three times slower than the standard NMS.”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data
    § 6.3. Test-time-augmentation ensemble · p. 5

    For the TTA ensemble of EfficientDetB7 on COCO, WBF achieved higher mAP than NMS, soft-NMS, and NMW.

    “| NMS | 0.5233 | 0.7129 | 0.5656 |”
    • Also stated in: § 6.3. Test-time-augmentation ensemble · p. 5 · plain
      “| Soft-NMS | 0.5210 | 0.7092 | 0.5633 |”
    • Also stated in: § 6.3. Test-time-augmentation ensemble · p. 5 · plain
      “| NMW | 0.5250 | 0.7138 | 0.5691 |”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given
    § 3. Weighted Boxes Fusion · p. 3

    In practice, the outcome for both variants of confidence rescaling did not differ significantly, with the first being slightly better.

    “In practice, the outcome for both variants did not differ significantly, with the first being slightly better.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation
    § 6.4.1 Ensemble of models for COCO Dataset · p. 5

    This result is a considerable improvement compared to the performance of individual models and gives the top third result on the MS COCO official leaderboard (as of 07/08/2020).

    “It is a considerable improvement compared to the performance of individual models and gives the top third result on the MS COCO official leaderboard [6] (as of 07/08/2020).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [6]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 7. Discussion · p. 7

    The 3D version of the WBF algorithm was successfully applied by the winners of the Waymo Detection challenge and the Lyft 3D Object Detection for Autonomous Vehicles challenge.

    “the winners from the first place of 3D Detection and Domain Adaptation track [13] and the second place from the 2D Object Detection track [11] used this method for combining predictions.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [13]

      Not checked in the free run.

    • cites [11]

      Not checked in the free run.

    • cites [19]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 2.2. Soft-NMS · p. 2

    Soft-NMS reduces confidences of proposals proportional to IoU instead of removing them, and improves performance over NMS on standard benchmarks for single models.

    “The soft-NMS method was proposed in [9] to reduce this problem. Instead of completely removing the detection proposals with high IoU and high confidence, it reduces the confidences of the proposals proportional to IoU value.”
    • Also stated in: § 2.2. Soft-NMS · p. 2 · boosted
      “Soft-NMS demonstrated noticeable improvements over traditional NMS on standard benchmark datasets, like PASCALVOCandMSCOCO[5]. It has shown to improve performance for single models [9].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [9]

      Not checked in the free run.

    • cites [5]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4. Datasets · p. 3

    The COCO dataset is one of the most popular benchmark datasets for object detection.

    “Microsoft COCO [5, 23], one of the most popular benchmark datasets for object detection.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [5]

      Not checked in the free run.

    • cites [23]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 3. Weighted Boxes Fusion · p. 3

    WBF uses all boxes, so it can fix cases where all boxes are predicted inaccurately by all models, whereas NMS/soft-NMS leave only one inaccurate box.

    “Both NMS and soft-NMS exclude some boxes, while WBF uses all boxes. Thus, it can fix cases where all boxes are predicted inaccurately by all models. This case is illustrated in Figure3. NMS/soft-NMS will leave only one inaccurate box, while WBF will fuse it using all predicted boxes.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning
    § 4.1. Open Images Dataset · p. 3

    As the test set labels are not available, we performed the ablation study on the validation set.

    “As the test set labels are not available, we performed the ablation study on the validation set.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The test set labels are not available.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 7. Discussion · p. 7

    WBF gives the best results on the Wheat detection dataset for boxes ensemble compared to other methods.

    “It was also reported that the WBF method gives the best results on the Wheat detection dataset for boxes ensemble comparing to other methods [32].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 7. Discussion · p. 7

    The first place winners from Lyft 3D Object Detection for Autonomous Vehicles reported that WBF gives better score for their ensemble (0.222 vs 0.220).

    “The first place winners from Lyft 3D Object Detection for Autonomous Vehicles [19] reported that WBF gives better score for their ensemble (0.222 vs 0.220) [48].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation
    § 6.4.1 Ensemble of models for COCO Dataset · p. 5

    These results can be reproduced using a benchmark published on GitHub [46].

    “These results can be reproduced using a benchmark published on GitHub [46].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [46]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 1. Introduction · p. 1

    Ensembles of models generalize better and usually yield more accurate results compared to a single model.

    “Combining predictions from different models generalizes better and usually yields more accurate results compared to a single model [27].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [27]

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning
    § 1. Introduction · p. 1

    NMS and soft-NMS only select boxes and cannot produce averaged localization of predictions combined from various models effectively.

    “Such methods work well on a single model. However, they only select the boxes and can not produce averaged localization of predictions combined from various models effectively.”
    • Also stated in: § 2.2. Soft-NMS · p. 2 · plain
      “However, both NMS and soft-NMS discard redundant boxes, and thus can not produce averaged localization predictions from different models effectively.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 6.2. An ensemble of two different models · p. 4

    These models have recently shown top results on the COCO database.

    “These models have recently shown top results on the COCO database [36].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning
    § 7. Discussion · p. 7

    The degradation from using WBF instead of NMS is due to the excessive number of low-scored wrong predictions given by the model output.

    “We think it is due to the excessive number of low scored wrong predictions given by the model output.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Most neural networks for Object Detection use NMS or Soft-NMS output to filter predicted boxes.

  • May combine two claims.

    Premisecitation
    § 3. Weighted Boxes Fusion · p. 3

    NMW does not change confidence scores; it uses IoU value to weight the boxes, and uses a box with the highest confidence to compare with, while WBF updates a fused box at each step.

    “The non-maximum weighted (NMW) method proposed in [49, 26] has a similar idea. However, the NMW method does not change confidence scores; it uses IoU value to weight the boxes. NMW uses a box with the highest confidence to compare with, while WBF updates a fused box at each step, and uses it to check the overlap with the next predicted boxes.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [49]

      Not checked in the free run.

    • cites [26]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductown data
    § 6.2. An ensemble of two different models · p. 4

    We used a grid search to find the optimal parameters for each of the four ensemble methods, which allowed achieving the best mAP for the given method.

    “We used a grid search to find the optimal parameters for each of the four methods, which allowed achieving the best mAP for the given method.”
    • Also stated in: § 6.3. Test-time-augmentation ensemble · p. 4 · plain
      “Then, we perform a grid search to find optimal parameters for all four ensemble methods.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 1. Introduction · p. 1

    The ensemble methods often get top places in machine learning competitions.

    “The ensemble methods often get top places in machine learning competitions, for example, see [12, 15, 34, 33, 42, 7].”
    Quote in the paperNot a substantial claim0 of 6 sources checked
    • cites [12]

      Not checked in the free run.

    • cites [15]

      Not checked in the free run.

    • cites [34]

      Not checked in the free run.

    • cites [33]

      Not checked in the free run.

    • cites [42]

      Not checked in the free run.

    • cites [7]

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning
    § 3. Weighted Boxes Fusion · p. 3

    NMW does not use information about how many models predict a given box in a cluster and therefore does not produce the best results for models' ensemble.

    “Also, NMW does not use information about how many models predict a given box in a cluster and, therefore, does not produce the best results for models' ensemble.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [49]

      Not checked in the free run.

    • cites [26]

      Not checked in the free run.

  • The quote only partly states this claim.

    Interpretationreasoning
    § 7. Discussion · p. 7

    NMS or soft-NMS produces more efficient filtering for low-scored wrong predictions than WBF.

    “The NMS or soft-NMS produces more efficient filtering for those predictions”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Most neural networks for Object Detection use NMS or Soft-NMS output to filter predicted boxes.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 7. Discussion · p. 7

    When there is a large number of overlapping boxes with different confidence scores, WBF gives worse results than NMS.

    “when it comes to a large number of overlapping boxes with different confidence scores, WBF gives worse results than NMS.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Most neural networks for Object Detection use NMS or Soft-NMS output to filter predicted boxes.

  • May combine two claims.

    Premisereasoning
    § 2.1. Non-maximum suppression (NMS) · p. 2

    In NMS, boxes are considered belonging to a single object if IoU is higher than a threshold, and setting this threshold is tricky because side-by-side objects may be eliminated, reducing precision.

    “Such errors reduce the precision of the model.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given
    § 6.4.1 Ensemble of models for COCO Dataset · p. 5

    The weights in the final ensemble and IoU threshold for the WBF methods were optimized using the validation set.

    “The weights in the final ensemble and IoU threshold for the WBF methods were optimized using the validation set.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation
    § 6.4.2 Ensemble of RetinaNet models for Open Images Dataset · p. 6

    These results can be reproduced using the benchmark published on GitHub.

    “These results can be reproduced using benchmark published on GitHub [46].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [46]

      Not checked in the free run.

29 of 29

  • The paper relies on this without support.

    The 3D version of boxes fusion was successfully applied by the winning teams of Waymo Open Dataset and Lyft 3D Object Detection for Autonomous Vehicles challenges.

    “The 3D version of boxes fusion was successfully applied by the winning teams of Waymo Open Dataset and Lyft 3D Object Detection for Autonomous Vehicles challenges.”

  • The paper relies on this without support.

    The Open Images dataset is the largest labeled dataset with object detection annotations to date.

    “We used the Open Images data set [20], the largest data set with object detection annotations to date”
    “To date, it is the largest labeled dataset with object detection annotations.”

  • The paper relies on this without support.

    The test set labels are not available.

    “As the test set labels are not available, we performed the ablation study on the validation set.”

  • The paper relies on this without support.

    Most neural networks for Object Detection use NMS or Soft-NMS output to filter predicted boxes.

    “Most neural networks for Object Detection use NMS or Soft-NMS output to filter predicted boxes.”
4 of 4

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • BB aggregation takes full consideration of object relationships and their spatial layout.

    “The advantage of this type of method is that it takes full consideration of object relationships and their spatial layout [118, 119].”

    From Object Detection in 20 Years: A Survey (2023) · cites this paper as [118]

    Section 3 (pp. 2–3) says WBF groups boxes by spatial overlap and fuses their coordinates, so the method accounts for the spatial positions of predicted boxes. It does not describe considering relationships between objects, and “full consideration” overstates the spatial processing it describes. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Inaccurate-box example cites the leaderboard figure

    Minor

    The text attributes the illustrated comparison of inaccurate predictions to Figure 3, but Figure 3 is captioned as a leaderboard. Figure 2 is captioned as the inaccurate-predictions comparison.

  • The individual-model columns in Tables 1 and 2 are mislabeled

    Minor

    The individual-model header labels both the second and third metric columns as mAP at 0.5 IoU, while the method header in the same table labels the third column as mAP at 0.75 IoU. The same conflicting headers appear in Table 2.

2 of 2

Cites

Flagged first, then by citations

3 of the 36 references couldn’t be matched to a known paper.

Cited by

Sorted by citations

Related

Authors
Solovyev, R., Wang, W., & Gabruseva, T.
Affiliations
  • Russian Academy of Sciences (RU)
  • National University of Singapore (SG)
  • Cork University Hospital (IE)
Corresponding author
Solovyev, Roman
ORCID
Venue
Image and Vision Computing
Publisher
Elsevier BV
Published
2021-03
Volume
107
Pages
104117
ISSN
0262-8856 (ISSN-L)
DOI
10.1016/j.imavis.2021.104117
OpenAlex
W3127743092
Semantic Scholar
2765484bdd3e5d8674917efc98145daa1a71e9fa
Cited by
543
References
36
Citations by year
2026: 108 · 2025: 98 · 2024: 92 · 2023: 130 · 2022: 76 · 2021: 36 · 2020: 3
FWCI
30.03
Type
journal article
Language
en
Open access
Licence
Locations
  • Image and Vision Computing (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Visual Attention and Saliency Detection
Keywords
object detection, model ensembling, confidence scores, 3D object detection, autonomous driving, Waymo Open Dataset

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review