Zero-Centered Fixed-Point Quantization With Iterative Retraining for Deep Convolutional Neural Network-Based Object Detectors
Open accessThe DOI matches its Crossref record; 49 of the 63 references it cites were found in a registry or an index, and 1 are not checked yet; 13 of the 27 sources its claims cite were checked for support.
Abstract
In the field of object detection, deep learning has greatly improved accuracy compared to previous algorithms and has been used widely in recent years. However, object detection using deep learning requires many hardware (HW) resources due to the huge computations for high performance, making it very difficult to run real-time on embedded platforms. Therefore, various compression methods have been studied to solve this problem. In particular, quantization methods greatly reduce the computational burden of deep learning by reducing the number of bits used for weights and activation functions in deep learning. However, most of these existing studies targeted only object classification and cannot be applied to object detection. Furthermore, most of the existing quantization studies are based on floating-point operations, which requires additional effort when implementing HW accelerators. This paper proposes an HW-friendly fixed-point-based quantization method that can also be applied to object detection. In the proposed method, the center of the weight distribution is adjusted to zero by subtracting the mean of weight parameters before quantization, and the retraining process is iteratively applied to minimize the accuracy drop caused by quantization. Furthermore, while applying the proposed method to object detection, performance degradation is minimized by considering the minimum and maximum values of weight parameters of deep learning networks. When applying the proposed quantization method to representative one-stage object detectors, You Only Look Once v3 and v4 (YOLOv3 and YOLOv4), detection accuracy similar to the original networks (i.e., YOLOv3 and YOLOv4) with a single-precision floating-point format (32-bit) is maintained despite expressing weights with only about 20% of the bits compared to a single-precision floating-point format in COCO dataset.
Verified 2026-11-19 · 1 reference not checked yet
Overview
CoChat Summary
## TL;DR
The paper proposes an HW-friendly fixed-point quantization method for deep CNNs that combines zero-centering of weights and iterative retraining, and extends it to object detectors with layer-wise bit allocation. On COCO, it reports near-original mAP for YOLOv3 and YOLOv4 while using only about 20% of the original weight bits.
## Key findings
- On CIFAR-10 with VGG-7, applying both zero-centered quantization and iterative retraining reduced accuracy drop to 0.1% at 4/4 bits; the paper reports 91.97% accuracy versus 92.07% for the original network.
- On YOLOv3 with COCO, the proposed method achieved 54.52 mAP at 6.58 average weight bits, compared with 54.64 mAP and 32 bits for the original model; the basic 8-bit quantized model was 52.60 mAP, while the proposed 8-bit model was 54.52 mAP.
- On YOLOv4 with COCO, the proposed method achieved 62.9 mAP at 8 bits versus 64.17 mAP for the original; with bit allocation, it reached 63.54 mAP at 5.99 average weight bits.
- The paper reports that iterative retraining reduced the portion of ambiguous weights from 23.33% to 8.73% in VGG-7, 22.95% to 18.01% in YOLOv3, and 20.06% to 14.65% in YOLOv4.
## Methods
- Proposed a fixed-point quantization scheme that subtracts the mean of layer weights to center them around zero, then quantizes and retrains iteratively while gradually increasing the quantization rate.
- Evaluated on CIFAR-10 with VGG-7 for image classification and COCO with YOLOv3/YOLOv4 for object detection; experiments were run on a Linux server with a GeForce RTX 2080Ti.
- For object detection, added a layer-wise bit allocation strategy using grid search to choose per-layer weight bitwidths that maximize mAP.
## Limitations
- The paper states that basic quantization without the proposed zero-centered and iterative retraining techniques causes severe performance degradation on object detection, even at 8 bits.
- The text makes evident that the method still has noticeable mAP loss at lower bitwidths for object detection: for example, YOLOv3 drops to 50.05 mAP at 6 bits, 46.95 at 5 bits, and 42.02 at 4 bits even with the proposed method.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures

















Verification
Claims the paper makes
5 Valid · 3 Not valid · 11 Not checked · 60 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Most recent quantization studies [21] [23] have focused on INT quantization with integer parameters and oatingpoint scale factors to prevent an accuracy decrease in networks that perform image classi cation.”
What the sources say
- Partly supportedcites [21] [21] S. Jung, C. Son, S. Lee, J. Son, J. J. Han, Y . Kwak, S. J. Hwang, and C. Choi, ''Learning to quantize deep networks by optimizing quantization intervals with task loss,'' in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , Jun. 2019, pp. 4350 4359. - sourceExperimental — accuracy not yet measured
The Related Work survey (Sec. 2, p. 2) and Sec. 3.1 (p. 3) do support part of the claim: recent quantization work is described as fixed-point/low-bit quantization of weights and activations that adds scaling factors (BWN's 'several scaling factors on the binary filters', XNOR-Net's 'filter-wise scaling factor', SYQ's 'pixel-wise or row-wise scaling') and is motivated by avoiding the 'severe accuracy degradation' of low-precision networks on image classification (Introduction, p. 1; Abstract, p. 1). But our copy nowhere states that MOST recent studies focus specifically on integer parameters with FLOATING-POINT scale factors — it never characterises the scale factors as floating-point, and much of what it surveys (BinaryConnect, INQ's logarithmic weights, DoReFa-Net's bitwidth study, Gupta's 16-bit fixed-point training, and its own EXP-Net training strategy) falls outside that description. The sweeping 'most recent studies' framing is therefore an overstatement of what this source reports. (LLM, conf=0.72)
- Could not verifycites [23] [23] S. Han, H. Mao, and W. J. Dally, ''Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding,'' 2015, arXiv:1510.00149 . [Online]. Available: http://arxiv.org/abs/1510.00149 - sourceExperimental — accuracy not yet measured
Our copy of Gaussian YOLOv3 contains no discussion of network quantization: it addresses localization uncertainty in object detection (Sec. 3.1 Gaussian modeling, Sec. 3.3 utilization of uncertainty) and accuracy/speed comparisons on KITTI and BDD (Sec. 4.2, Tables 2-4). Searches for 'quantization / INT quantization / fixed-point integer parameters / floating-point scale factors / image classification' returned only semantic neighbours about detection mAP, FLOPs and Gaussian bbox modeling, with no passage on quantized parameters or scale factors. The claim's point is therefore not addressed by this source, and no alternative work is named in our copy to credit it to. (LLM, conf=0.75)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“The proposed technique can be easily applied to various object detectors including YOLOv3 [10] and YOLOv4 [11].”
What the sources say
- Supportedcites [10] [10] J. Redmon and A. Farhadi, ''YOLOv3: An incremental improvement,'' 2018, arXiv:1804.02767 . [Online]. Available: http://arxiv.org/abs/1804.02767 - source
The citation is attached to YOLOv3 as an example of an object detector. The paper identifies and describes YOLOv3 as a detector in “3. How We Do” (p. 3); the citing authors’ assertion about applying their proposed technique is not attributed to this source. (LLM, conf=0.91)
- Supportedcites [11] [11] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, ''YOLOv4: Optimal speed and accuracy of object detection,'' 2020, arXiv:2004.10934 . [Online]. Available: http://arxiv.org/abs/2004.10934 - source
The citation is attached to YOLOv3 as an example of an object detector, not to the citing authors’ claim that their proposed technique is easily applicable. Section 2.3 (p. 2) describes YOLOv3 predicting bounding boxes across three scales, confirming that identification. (LLM, conf=0.94)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“However, despite using 1 bit more for the activation parameter than the proposed method, the accuracy drop of Miyashita et al. [39] is greater by 0.21% than that of the proposed method.”
What the sources say
- Could not verifycites [39] [39] D. Miyashita, E. H. Lee, and B. Murmann, ''Convolutional neural networks using logarithmic data representation,'' 2016, arXiv:1603.01025 . [Online]. Available: https://arxiv.org/abs/1603.01025 - sourceExperimental — accuracy not yet measured
Our copy of Nayak et al. (2019) contains no comparison with Miyashita et al.: searches for 'Miyashita', for the 0.21% accuracy-drop differential, and for activation-bit comparisons returned nothing. The closest passages are IV.a (p. Table 1), which only says ResNet-18 CIFAR-10 results are shown 'at byte boundary for comparison purposes with published results' without naming any work or giving a differential, and III/IV.b, which describe quantization of model parameters only (convolution kernels, batch-norm and dense layers) with accuracy drops of '~4% up to 6-bit' — no activation bit-widths are reported. Thus neither the 0.21% figure nor the 1-bit activation difference can be checked against this source; nothing in our copy contradicts it either. (LLM, conf=0.68)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“In particular, through the comparison result with Gaussian YOLOv3-tiny [13], which enhances the accuracy of a small-scale network, it can be seen that effectively reducing the size of a large-scale network is more ef cient than improving the performance of a smallscale network.”
What the sources say
- Could not verifycites [13] [13] J. Choi, D. Chun, H.-J. Lee, and H. Kim, ''Uncertainty-based object detector for autonomous driving embedded platforms,'' in Proc. 2nd IEEE Int. Conf. Artif. Intell. Circuits Syst. (AICAS) , Aug. 2020, pp. 16 20. - sourceExperimental — accuracy not yet measured
The Methodology and Selection of architecture sections (p. 5) discuss balancing speed, accuracy, and model size, but do not compare shrinking a large network with improving a small one. Searches for Gaussian YOLOv3-tiny and the claimed efficiency comparison found no passage addressing that conclusion. (LLM, conf=0.76)
plainstates a causeSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The results illustrate that the portion of the ambiguous weight parameters is signi cantly reduced during retraining from 23.33%, 22.95%, and 20.06% to 8.73%, 18.01%, and 14.65% in VGG-7, YOLOv3, and YOLOv4, respectively”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
30 more claims
Create a free account to see all 79 claims and to verify the 9 references still open.
- It may state two separate claims as one. 37
- The quote only partly states this claim. 4
- Probably not a substantial claim. 11
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
Very Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×
“Table 2 illustrates the accuracy results of the existing VGG-7 network ( i.e. , 32-bit full decision) [2] and the quantized VGG-7 networks with the various options in the second column.”
BackgroundImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 1×
“detection and classification increases dramatically [1]–[4].”
ImageNet: A large-scale hierarchical image database
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333 · cited 1×
“Especially, Jung et al. [21] found that a reduction in accuracy does not occur when using only four bits for both weight and activation parameters, even for the ImageNet dataset [32].”
BackgroundYou Only Look Once: Unified, Real-Time Object Detection
Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.91 · Cited by 53,803 · cited 2×
“In other words, YOLOv3 is a fully-convolutional network composed of only 1 × 1 and 3 × 3 size convolution filters and illustrates outstanding accuracy for objects of various sizes while maintaining a high processing speed similar to YOLO [33] and YOLOv2 [34].”
MethodResultMicrosoft COCO: Common Objects in Context
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10602-1_48 · Cited by 45,460 · cited 1×
“5 illustrates the medians of maximum and minimum weight parameters for each layer when training the VGG-7 network with Cifar-10 dataset [35] and the YOLOv3/v4 networks with COCO dataset [36].”
Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014) · 2014 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2014.81 · Cited by 32,684 · cited 1×
“In particular, for DNNbased object detection, classification, and localization are performed simultaneously, which requires vast computation [8]–[11].”
Background(2016) · Cited by 31,590
Learning Multiple Layers of Features from Tiny Images
Krizhevsky, A. (2024) · doi:10.57702/zp44cu3g · Cited by 23,395
SSD: Single Shot MultiBox Detector
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., & Berg, A. C. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-46448-0_2 · Cited by 22,128
Cited by
A Survey on Adversarial Attack in the Age of Artificial Intelligence
Kong, Z., Xue, J., Wang, Y., Huang, L., Niu, Z., & Li, F. (2021) · Wireless Communications and Mobile Computing · doi:10.1155/2021/4907754 · Cited by 98
“Then, the neural network model is attacked, and the process is repeated [81].”
Toward AI-Assisted UAV for Human Detection in Search and Rescue Missions
Rizk, M., Slim, F., & Charara, J. (2021) · International Conference on Decision Aid Sciences and Applications (DASA) · doi:10.1109/dasa53625.2021.9682412 · Cited by 58
“1 shows the network architecture of YOLOv3 [16].”
BackgroundKim, N. J., & Kim, H. (2022) · IEEE Transactions on Multimedia · doi:10.1109/tmm.2022.3189496 · Cited by 56
Self-citationA survey of FPGA and ASIC designs for transformer inference acceleration and optimization
Kang, B. J., Lee, H. I., Yoon, S. K., Kim, Y. C., Jeong, S. B., Jeong, S. B., & Kim, H. (2024) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2024.103247 · Cited by 63
HAWQ-V3: Dyadic Neural Network Quantization
Yao, Z., Dong, Z., Zheng, Z., Gholaminejad, A., Yu, J., Tan, E., Wang, L., Huang, Q., Wang, Y., Mahoney, M. W., et al. (2021) · International Conference on Machine Learning · Cited by 281
Metadata
- Authors
- Kim, S., & Kim, H.
- Affiliations
- Seoul National University of Science and Technology (KR)
- Venue
- IEEE Access
- Published
- 2021
- Volume
- 9
- Pages
- 20828-20839
- ISSN
- 2169-3536 (ISSN-L)
- OpenAlex
- W3122390253
- Cited by
- 71
- References
- 63
- Citations by year
- 2026: 3 · 2025: 8 · 2024: 22 · 2023: 11 · 2022: 15 · 2021: 11
- FWCI
- 4.52
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- IEEE Access (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Institute of Information & Communications Technology Planning & Evaluation (IITP) Grant funded by the Korean Government (MSIT) (2020-0-01305)
- Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2019R1A6A1A03032119)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Computational Theory and Mathematics › Numerical Methods and Algorithms
- Keywords
- fixed-point quantization, object detection, model compression, YOLOv3, YOLOv4, weight distribution, embedded hardware acceleration, low-bit precision, accuracy degradation
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).
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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Zero-Centered Fixed-Point Quantization With Iterative Retraining for Deep Convolutional Neural Network-Based Object Detectors. https://app.cochat.ai/papers/10.1109/access.2021.3054879
- The paper
- Kim, S., & Kim, H. (2021). Zero-Centered Fixed-Point Quantization With Iterative Retraining for Deep Convolutional Neural Network-Based Object Detectors. IEEE Access. https://doi.org/10.1109/access.2021.3054879
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.