Pith. sign in

REVIEW 4 major objections 4 minor 52 references

QATMA: Quantization-Aware Training with Multimodal Alignment for Open-Vocabulary Object Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Extreme 4-bit quantization of open-vocabulary detectors is recoverable if the model is quantized stage-by-stage and distilled through text-anchored pairwise similarity matrices.

desk verdict A plausible, well-specified method for 4-bit OVOD quantization, but the empirical support is a single stripped baseline with no code or seeds — worth refereeing, not worth trusting yet. read the letter →

arxiv 2603.05964 v3 pith:2Z6RSOCP submitted 2026-03-06 cs.CV

Jinyeong Park , Donghwa Kang , Seunghwan An , Insoo Kim , Brent ByungHoon Kang , Hyeongboo Baek , Jibum Kim This is my paper · ORCID
classification cs.CV
keywords open-vocabularyobjectdetectionquantization-awaretraininglow-bitquantizationknowledgedistillationvision-languagealignmentrelationalstructurecurriculumlearningtext-anchoredsimilarity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that extreme low-bit quantization (4-4-8 bit) breaks open-vocabulary detectors not merely by lowering accuracy but by distorting the joint vision-language embedding space: the similarity between region and text embeddings, and the pairwise relational structure among regions of a category. It argues that standard QAT trained with only the task loss cannot repair these distortions, and proposes CR-QAT, which quantizes the backbone first with the neck-head frozen, then quantizes the neck-head, while distilling the full-precision teacher's text-anchored pairwise similarity matrices into the student. On LVIS and COCO zero-shot benchmarks, CR-QAT beats the QAT baseline by up to +4.2 and +7.6 AP, i.e., relative gains of 38.9% and 40.9%, which would make 4-bit open-vocabulary detectors practical for embedded deployment. The work positions itself as the first to tackle extreme low-bit quantization for open-vocabulary detectors.

What carries the argument

Text-anchored pairwise similarity matrices: for each text query c, the method stacks the text embedding t_c with N_c region embeddings assigned by the teacher's task-aligned label assignment (TAL), row-wise L2-normalizes, and forms S_c = X_c X_c^T; the first row/column encodes region-text alignment, the rest encodes region-region relationships, and the student is trained to match the teacher's S_c via Smooth L1. The companion mechanism is curriculum QAT (CQAT): the model is split into functional units and quantized sequentially, here in two stages — backbone first with the neck-head frozen, then neck-head — so that unquantized modules isolate errors and give stable gradients. The two mechani

What would settle it

Hold CR-QAT fixed but replace the teacher's TAL-assigned positive regions per text query with a fixed IoU-based selection on YOLO-World-X; a drop from +4.2 AP over QAT to near zero would confirm that reliable TAL anchors carry the result.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the multi-dimensional relational structure of an OVOD model — the cosine similarities between a text query and the region embeddings assigned to it, and the pairwise similarities among those regions — is the load-bearing quantity that low-bit quantization destroys, and that this structure can be explicitly transferred from a full-precision teacher by building text-anchored pairwise similarity matrices and minimizing their discrepancy. The authors also find that curriculum staging is what makes this distillation effective: applying KD without the stage-by-stage quantization yields almost no gain (+0.4 AP), the curriculum without KD yield

Load-bearing premise

The method's benefit rests on the teacher's TAL-positive regions being reliable anchors for each text query, since noisy assignments would consolidate the wrong relational structure during distillation.

Editorial extensions

If this is right

  • Under 4-4-8 (Ch-T-H) quantization, CR-QAT adds +2.5, +3.4, and +4.2 AP over QAT on YOLO-World-M/L/X for LVIS zero-shot, and +4.1, +6.8, and +7.6 AP on COCO.
  • Gains concentrate on rare categories: APr improves by +5.2 to +5.7 AP on LVIS-L/X, which the paper takes as evidence that damaged vision-language alignment is being restored.
  • The benefit persists across bit-widths and granularities: 4-5-8 (Ch-T-H) yields +0.6 AP and 3-3-8 (Ch-Ch-H) yields +1.3 AP over QAT.
  • Embedding-space relational distortion correlates with confidence-space distortion (Spearman ρ=0.76), so preserving inter-region relationships in the embedding carries through to final predictions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The method is claimed to be architecture-agnostic, but only YOLO-World is tested; a natural test is to apply the same two-stage curriculum and text-anchored distillation to transformer-based OVOD detectors with online cross-modal fusion, where the neck-head grouping may need to be redrawn.
  • If the teacher's TAL assignment is itself noisy on rare categories, TRKD could reinforce false region-text relationships; a robust variant might weight the matrix loss by assignment confidence or use multiple positive sets.
  • The paper's observation that KD alone fails without the curriculum suggests that for even lower bit-widths (3-bit weight/activation), the curriculum may need more than two stages, or a different freeze schedule; this is an untested extrapolation.
  • A deployment-oriented spin-off: instead of recomputing S_c every iteration, cache the teacher's pairwise matrices offline as fixed soft targets, which could cut the distillation memory footprint while retaining most of the relational signal.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes CR-QAT (also called QATMA in the abstract), a quantization-aware training framework for open-vocabulary object detection. It consists of curriculum QAT (CQAT), which progressively quantizes the backbone then the neck-head, and text-centric relational knowledge distillation (TRKD), which transfers teacher-student region-text and region-region pairwise similarity matrices. Experiments on YOLO-World with 4-4-8-bit weight-activation-attention quantization report consistent AP gains over a plain LSQ QAT baseline on LVIS and COCO zero-shot benchmarks, with relative improvements up to 38.9% and 40.9%.

Significance. If the empirical claims hold, this is a useful first systematic study of extreme low-bit quantization for OVOD models. The method is described in enough detail to reproduce (equations for quantization, curriculum objective, TRKD loss, and hyperparameters), and the ablations are internally coherent. However, the evidence base is currently too narrow: the only QAT comparator is a self-implemented plain-LSQ baseline, no code or seeds are provided, and all results are single runs without variance. The abstract/full-text naming and AP-gain mismatch (QATMA vs CR-QAT; 4.3 vs 4.2) further reduces confidence. The central claim that CR-QAT 'consistently outperforms existing QAT baselines' is therefore not yet established to the standard expected for a venue that publishes comparative empirical claims.

major comments (4)
  1. [§5.2, Tables 1–2] The only QAT comparator is the authors' own LSQ-based baseline trained with task loss only. The abstract and §1 claim 'consistently outperforms existing QAT baselines', but no existing low-bit QAT method (e.g., LSQ+ [2], QKD [23], Q-DETR [43], or a feature-distillation baseline) is compared. Because CR-QAT adds a curriculum plus two KD losses, the reported +2.5 to +7.6 AP gains could in principle come from generic KD or the curriculum alone. Please add comparisons with at least one published QAT/KD method adapted to YOLO-World, and report the plain-LSQ baseline with the same training budget and also with feature distillation only.
  2. [Abstract vs §5.2/Table 1] The abstract describes the method as 'QATMA' and reports an AP gain of 'up to 4.3' on LVIS; the full text uses 'CR-QAT' and Table 1 reports +4.2 AP for YOLO-World-X. This is a material inconsistency between the two versions of the paper. It must be reconciled, and the final claim should match the numbers in the tables.
  3. [§5.4, Tables 3–6] All experiments are single runs with no seeds, error bars, or confidence intervals. Several decisive comparisons are numerically small: Table 4 shows 13.1 vs 13.7 vs 14.3; Table 6 shows 22.0 vs 21.4 for 4-5-8. Without multiple seeds and variance, the ablations and granularity conclusions are not statistically grounded. Please report mean±std over at least three seeds, and release code or a reproducible configuration.
  4. [§1, §4.1, §5] The paper claims an 'architecture-agnostic' framework and states in §4.1 that the two-stage curriculum applies to 'widely-adopted OVOD architectures (e.g., GLIP, YOLO-World)', yet experiments cover only YOLO-World. This is an overclaim relative to the evidence. Either add a demonstration on a second architecture (e.g., GLIP-T or Grounding DINO) or restrict the claims to YOLO-World and motivate transferability only as future work.
minor comments (4)
  1. [§4.2, Eq. (10)] Clarify whether the same N_c region embeddings are used for teacher and student, and how regions are selected when the student's TAL assignment differs from the teacher's. The current text says positive regions are determined by the teacher's TAL; if the student uses the same assignment, this should be stated explicitly.
  2. [§5.1] It is unclear which layers are 'first and last layers' in the YOLO-World instantiation (backbone stem? final head?). Also, state whether the text encoder is kept in FP32 for both teacher and student during training, since it is excluded from quantization.
  3. [§5.3] The sentence 'distilling multi-scale feature pyramids proves unstable under low-bit settings [50]' has a stray superscript '4' after the citation. Also, footnote 4 says three-stage with neck distillation yields 13.3, which should be incorporated into Table 5 for completeness.
  4. [§5.4, Fig. 3] The embedding-relation MAE in Fig. 3 is essentially the TRKD objective, so the high correlation with confidence-relation MAE is informative but not an independent confirmation of the method's benefit. This should be noted in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the method's efficacy is evaluated on held-out AP benchmarks rather than its own training objective.

full rationale

CR-QAT's derivation chain—motivating quantization-induced loss of region-text and region-region structure, proposing curriculum QAT (CQAT) and text-centric relational KD (TRKD), and evaluating on LVIS/COCO zero-shot AP—does not reduce to its inputs by construction. The reported AP improvements (Tables 1-2) are measured on external benchmarks (LVIS miniVal fixed AP, COCO val2017) against a QAT baseline, not on the training loss itself. The ablations (Tables 3-5) decompose the contributions of curriculum and KD components and would be meaningless if the result were definitionally forced. The embedding-distortion analyses (Fig. 1, Fig. 3) do track the TRKD objective (Eq. 10), since TRKD directly minimizes the discrepancy between teacher and student pairwise similarity matrices; thus those plots are illustrative of the loss behavior rather than independent confirmation. However, that is not circularity—the central claim of improved AP is independently evaluated. No self-citations are used as load-bearing evidence; references to error accumulation [47] and feature-distillation instability [50] are external prior works. The abstract/full-text naming and AP-gain discrepancy (QATMA vs CR-QAT, 4.3 vs 4.2) is an internal inconsistency that affects reproducibility confidence, but it does not constitute a circular derivation. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard QAT and distillation heuristics plus several hand-set hyperparameters. There are no invented physical entities. The main unquantified assumptions concern the reliability of teacher-generated supervision signals and the representativeness of the single YOLO-World benchmark.

free parameters (3)
  • KD loss weight lambda_i = 6.0
    Used for both feature distillation and TRKD (Sec 5.1); chosen by hand; no sensitivity analysis reported.
  • Curriculum stage data split = 1/3 backbone, 2/3 neck-head
    Stage 1 uses first 1/3 of data, stage 2 uses remaining 2/3; no ablation of the split ratio is provided.
  • LSQ learning-rate ratio = 0.1 x base learning rate
    Quantization scale/offset learning rate is set to 0.1x the base rate (Sec 5.1); a hand-set hyperparameter.
assumptions (4)
  • domain assumption Uniform quantization with the straight-through estimator gives usable gradients for training quantized networks.
    Invoked in Sec 3.1; standard QAT assumption but heuristic and not validated for YOLO-World in this paper.
  • domain assumption The FP32 teacher's text embeddings and TAL positive-region assignments provide reliable relational knowledge to transfer.
    Core to TRKD (Sec 4.2, Sec 5.1); if the teacher's alignments are noisy, the distillation targets are flawed.
  • domain assumption Objects365v2 + GQA training with YOLO-World's pretraining pipeline is a valid zero-shot proxy for LVIS/COCO evaluation.
    Sec 5.1 relies on the YOLO-World protocol; category overlap and zero-shot validity are not re-examined.
  • domain assumption Early-layer quantization error accumulation is the dominant failure mode, motivating stage-by-stage optimization.
    Sec 4.1 motivates CQAT by citing [47]; this is not directly measured for YOLO-World in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QATMA: Quantization-Aware Training with Multimodal Alignment for Open-Vocabulary Object Detection." pith.science (2026). https://pith.science/paper/2Z6RSOCP

@misc{pith2026260305964,
  author       = {Pith},
  title        = {Pith review of: QATMA: Quantization-Aware Training with Multimodal Alignment for Open-Vocabulary Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2Z6RSOCP}},
  note         = {Machine review of arXiv:2603.05964}
}
read the original abstract

Quantizing open-vocabulary object detection (OVOD) models reduces their memory and computational costs, but extremely low-bit quantization severely degrades both cross-modal (region-text) and intra-modal (region-region) alignments. This multimodal degradation is a unique challenge that prior quantization methods for closed-vocabulary detectors fail to resolve. To overcome this, we propose Quantization-Aware Training with Multimodal Alignment (QATMA), the first multimodal-aware and architecture-agnostic QAT framework tailored for OVOD. QATMA integrates two key components: (i) Curriculum QAT, which partitions the detector by functional roles and progressively expands the quantization scope to suppress error accumulation and ensure stable optimization; and (ii) Text-anchored Pairwise Similarity Distillation, which transfers both region-text and region-region alignments from a full-precision teacher model via pairwise cosine similarities in the joint embedding space. Experimental results on LVIS and COCO zero-shot benchmarks demonstrate that QATMA significantly outperforms existing QAT baselines under extremely low-bit settings, achieving gains of up to 4.3 and 7.6 AP, respectively.

Figures

Figures reproduced from arXiv: 2603.05964 by the authors.

Figure 1
Figure 1. Impact of 4-bit quantization on YOLO-World [7] with the Objects365v2 [40] dataset. (a) Confidence scores derived from the similarity between each region embed￾ding and the text embedding of category “Lamp” (P¯: mean over positive regions). (b) Pairwise cosine similarity matrix among positive region embeddings within the same category (r: Pearson correlation with FP32). (c) Quantitative comparison of embed￾ding disto… view at source ↗
Figure 2
Figure 2. Overview of the proposed CR-QAT framework. (Red) blocks denote quantized and learnable modules, and (blue) blocks denote full-precision and frozen modules. (a) Stage 1: the backbone (M1) is quantized with Lfeat supervision from the full￾precision teacher, while the neck-head remains frozen for error isolation. (b) Stage 2: the neck-head (M2) is additionally quantized, supervised by both Lfeat and LTRKD. (c) Feature … view at source ↗
Figure 3
Figure 3. Correlation between embedding￾level and confidence-level inter-region re￾lation distortion relative to FP32. Each point represents an (image, category) group with ≥10 anchors. (Large markers) denote the mean over all groups. (ρ) de￾notes the Spearman correlation. 5.4 Analysis Effect of Quantization Granularity [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on YOLO-World-L (4-4-8, Ch-T-H). (Top) Detection results. (Bottom) Inter-region similarity heatmap of average pairwise cosine similarity among same-class anchors. QAT distorts both detection and similarity patterns of FP32, whereas CR-QAT restore…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 6 linked inside Pith

  1. [2]

    In: CVPRW

    Bhalgat, Y., Lee, J., Nagel, M., Blankevoort, T., Kwak, N.: LSQ+: Improving low- bit quantization through learnable offsets and better initialization. In: CVPRW. pp. 696–697 (2020)

    2020

  2. [23]

    arXiv preprint arXiv:1911.12491 (2019)

    Kim,J.,Bhalgat,Y.,Lee,J.,Patel,C.,Kwak,N.:QKD:Quantization-awareknowl- edge distillation. arXiv preprint arXiv:1911.12491 (2019)

    arXiv 1911

  3. [43]

    In: CVPR

    Xu, S., Li, Y., Lin, M., Gao, P., Guo, G., Lü, J., Zhang, B.: Q-DETR: An efficient low-bit quantized detection transformer. In: CVPR. pp. 3842–3851 (2023)

    2023

  4. [1]

    arXiv preprint arXiv:1308.3432 (2013)

    Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradi- ents through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)

    arXiv 2013

  5. [3]

    In: NeurIPS

    Cao, W., Zhang, Y., Gao, J., Cheng, A., Cheng, K., Cheng, J.: PKD: General distillation framework for object detectors via pearson correlation coefficient. In: NeurIPS. pp. 15394–15406 (2022)

    2022

  6. [4]

    In: ECCV

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End- to-end object detection with transformers. In: ECCV. pp. 213–229 (2020)

    2020

  7. [5]

    In: NeurIPS

    Chen, G., Choi, W., Yu, X., Han, T., Chandraker, M.: Learning efficient object detection models with knowledge distillation. In: NeurIPS. vol. 30 (2017)

    2017

  8. [6]

    In: CVPR

    Chen, P., Liu, J., Zhuang, B., Tan, M., Shen, C.: AQD: Towards accurate quantized object detection. In: CVPR. pp. 104–113 (2021)

    2021

Show all 52 references
  1. [7]

    In: CVPR

    Cheng, T., Song, L., Ge, Y., Liu, W., Wang, X., Shan, Y.: YOLO-World: Real-time open-vocabulary object detection. In: CVPR. pp. 16901–16911 (2024)

    2024

  2. [8]

    arXiv preprint arXiv:1805.06085 (2018)

    Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrish- nan, K.: PACT: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)

    2018 arXiv

  3. [9]

    In: ICCV Workshops

    Choukroun, Y., Kravchik, E., Yang, F., Kisilev, P.: Low-bit quantization of neural networks for efficient inference. In: ICCV Workshops. pp. 3009–3018 (2019)

    2019

  4. [10]

    In: ICLR (2020)

    Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: ICLR (2020)

    2020

  5. [11]

    In: ICCV

    Feng, C., Zhong, Y., Gao, Y., Scott, M.R., Huang, W.: TOOD: Task-aligned one- stage object detection. In: ICCV. pp. 3490–3499 (2021)

    2021

  6. [12]

    In: CVPR

    Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accu- rate object detection and semantic segmentation. In: CVPR. pp. 580–587 (2014)

    2014

  7. [13]

    In: ICLR (2022)

    Gu, X., Lin, T., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022)

    2022

  8. [14]

    In: CVPR

    Gupta, A., Dollár, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356–5364 (2019)

    2019

  9. [15]

    In: WACV

    Gupta, K., Asthana, A.: Reducing the side-effects of oscillations in training of quantized YOLO networks. In: WACV. pp. 2452–2461 (2024)

    2024

  10. [16]

    In: NeurIPS

    Han, S., Pool, J., Tran, J., Dally, W.: Learning both weights and connections for efficient neural network. In: NeurIPS. vol. 28 (2015)

    2015

  11. [17]

    In: ICCV

    He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: ICCV. pp. 1389–1397 (2017)

    2017

  12. [18]

    arXiv preprint arXiv:1503.02531 (2015)

    Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  13. [19]

    In: CVPR

    Hudson, D.A., Manning, C.D.: GQA: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR. pp. 6700–6709 (2019)

    2019

  14. [20]

    In: CVPR

    Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko,D.:Quantizationandtrainingofneuralnetworksforefficientinteger- arithmetic-only inference. In: CVPR. pp. 2704–2713 (2018)

    2018

  15. [21]

    In: ICML

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904–4916 (2021) 16 J. Park et al

    2021

  16. [22]

    Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLOv8.https://github.com/ ultralytics/ultralytics(2023)

    2023

  17. [24]

    In: CVPR

    Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J., Chang, K., Gao, J.: Grounded language-image pre-training. In: CVPR. pp. 10955–10965 (2022)

    2022

  18. [25]

    In: CVPR

    Li, R., Wang, Y., Liang, F., Qin, H., Yan, J., Fan, R.: Fully quantized network for object detection. In: CVPR. pp. 2810–2819 (2019)

    2019

  19. [26]

    In: ICCV

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection. In: ICCV. pp. 2999–3007 (2017)

    2017

  20. [27]

    In: ECCV

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV. pp. 740–755 (2014)

    2014

  21. [28]

    In: ECCV

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., Zhu, J., Zhang, L.: Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In: ECCV. pp. 38–55 (2024)

    2024

  22. [29]

    In: ECCV

    Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: SSD: Single shot multibox detector. In: ECCV. pp. 21–37 (2016)

    2016

  23. [30]

    In: ICCV

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012–10022 (2021)

    2021

  24. [31]

    In: ECCV

    Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Doso- vitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV. pp. 728–755 (2022)

    2022

  25. [32]

    In: ICML

    Nagel, M., Amjad, R.A., Van Baalen, M., Louizos, C., Blankevoort, T.: Up or down? adaptive rounding for post-training quantization. In: ICML. pp. 7197–7206 (2020)

    2020

  26. [33]

    In: ICCV

    Nagel, M., Baalen, M.v., Blankevoort, T., Welling, M.: Data-free quantization through weight equalization and bias correction. In: ICCV. pp. 1325–1334 (2019)

    2019

  27. [34]

    In: ICML

    Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscilla- tions in quantization-aware training. In: ICML. pp. 16318–16330 (2022)

    2022

  28. [35]

    In: ICLR (2018)

    Polino, A., Pascanu, R., Alistarh, D.: Model compression via distillation and quan- tization. In: ICLR (2018)

    2018

  29. [36]

    In: ICML

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763 (2021)

    2021

  30. [37]

    In: CVPR

    Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: CVPR. pp. 779–788 (2016)

    2016

  31. [38]

    arXiv preprint arXiv:1804.02767 (2018)

    Redmon, J., Farhadi, A.: YOLOv3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)

    2018 arXiv

  32. [39]

    In: NeurIPS

    Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: NeurIPS. pp. 91–99 (2015)

    2015

  33. [40]

    In: ICCV

    Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: ICCV. pp. 8430–8439 (2019)

    2019

  34. [41]

    In: ICCV

    Tian, Z., Shen, C., Chen, H., He, T.: FCOS: Fully convolutional one-stage object detection. In: ICCV. pp. 9627–9636 (2019)

    2019

  35. [42]

    In: CVPR

    Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR. pp. 7464–7475 (2023) CR-QAT for Open-Vocabulary Object Detection 17

    2023

  36. [44]

    In: CVPR

    Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR. pp. 14393–14402 (2021)

    2021

  37. [45]

    In: ICLR (2023)

    Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023)

    2023

  38. [46]

    In: NeurIPS

    Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: GLIPv2: Unifying localization and vision-language under- standing. In: NeurIPS. pp. 36067–36080 (2022)

    2022

  39. [47]

    In: ICCV

    Zhang, Y., Dong, Z., Yang, H., Lu, M., Tseng, C.C., Du, Y., Keutzer, K., Du, L., Zhang, S.: QD-BEV: Quantization-aware view-guided distillation for multi-view 3D object detection. In: ICCV. pp. 3825–3835 (2023)

    2023

  40. [48]

    In: CVPR

    Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: RegionCLIP: Region-based language-image pretraining. In: CVPR. pp. 16793–16803 (2022)

    2022

  41. [49]

    arXiv preprint arXiv:1606.06160 (2016)

    Zhou, S., Ni, Z., Zhou, X., Wen, H., Wu, Y., Zou, Y.: DoReFa-Net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016)

    2016 arXiv

  42. [50]

    In: AAAI

    Zhu, K., He, Y.Y., Wu, J.: Quantized feature distillation for network quantization. In: AAAI. vol. 37, pp. 11452–11460 (2023)

    2023

  43. [51]

    In: ICLR (2021)

    Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. In: ICLR (2021)

    2021

  44. [52]

    In: CVPR

    Zhuang, B., Liu, L., Tan, M., Shen, C., Reid, I.: Training quantized neural networks with a full-precision auxiliary module. In: CVPR. pp. 1488–1497 (2020)

    2020

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.