Pith. sign in

REVIEW 3 major objections 6 minor 31 references

A lightweight DETR can run its entire forward pass in integer arithmetic, cutting size by about 3.6× and compute by more than 10× with only moderate detection loss.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A fully integer-only lightweight DETR with split-scale projectors, sign-dependent ShiftGELU, and constrained Shiftmax cuts size ~3.6× and compute >10× at moderate mAP cost.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid systems paper that actually closes fully-integer lightweight DETR; the bilinear→NN swap is a real but secondary attribution soft spot, not a collapse of the claim. the 3 major comments →

arxiv 2607.24981 v1 pith:QAMFZALF submitted 2026-07-27 cs.CV

Enabling Fully Integer-Only Inference for Lightweight Detection Transformers

Thanh Cong Le , Michal Szczepanski , Martyna Poreba This is my paper
classification cs.CV
keywords ViTObject DetectionDETRQuantizationInteger-only inferenceLightweight transformersDeformable attention
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Lightweight transformer detectors are accurate enough to rival CNNs, but Softmax, GELU, deformable attention, and multi-scale fusion still force floating-point math that edge NPUs and microcontrollers cannot run efficiently. This paper builds I-LW-DETR so that every forward-pass operation, including the transformer nonlinearities, stays in integer arithmetic. Three redesigns carry the result: a split projector convolution that keeps a separate activation scale per feature branch, a sign-dependent GELU that never forces activations into the high-error region of the integer exponential, and a calibrated Shiftmax that keeps Softmax normalization inside 32-bit range. Across Tiny, Small, and Medium models the pipeline loses only a few mAP points after quantization-aware fine-tuning while shrinking the model roughly 3.6× and the bit-operation count by more than an order of magnitude. A sympathetic reader cares because this removes the last floating-point barrier that has kept deformable DETRs off integer-only edge silicon.

Core claim

I-LW-DETR is the first fully integer-only lightweight DETR: after quantization-ready architectural changes and three integer approximations, every operation in the forward pass—including deformable attention, multi-scale projection, Softmax, GELU, and LayerNorm—executes in integer arithmetic, and the resulting models retain competitive COCO and VisDrone accuracy at roughly 3.6× smaller size and more than 10× lower bit-operation cost.

What carries the argument

Three integer blocks that jointly close the gap: (1) scale-preserving split convolution, which gives each multi-scale projector branch its own activation scale before a dyadic rescale-and-add; (2) SD-ShiftGELU, a sign-dependent GELU that evaluates the exponential on each element independently so no vector-max push occurs; (3) constrained Shiftmax, which calibrates a per-block denominator shift so the Softmax sum stays representable in 32-bit integers.

Load-bearing premise

Replacing bilinear sampling with nearest-neighbor and sinusoidal box encodings with a learned linear projection, plus a fixed calibration margin on the Softmax denominator, still keeps enough localization and attention fidelity for practical detection, including small objects.

What would settle it

Measure COCO and VisDrone mAP of the full integer pipeline against the floating-point QR-LW-DETR baseline after the same QAT schedule; if the drop exceeds the reported 4–5 mAP on COCO (or the larger small-object drop on VisDrone) once nearest-neighbor and the calibrated Shiftmax are locked in, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Lightweight deformable DETRs become deployable on integer-only NPUs and microcontrollers without leaving Softmax, GELU, or LayerNorm in floating point.
  • Model size falls by roughly 3.6× and bit-operations by 12–14× across Tiny/Small/Medium scales while QAT recovers most of the PTQ accuracy loss.
  • The same three integer blocks transfer from natural images (COCO) to dense aerial small-object scenes (VisDrone), though small-object AP remains the hardest regime.
  • Operator-level mapping shows the graph is expressible with standard integer primitives (convolutions, matmuls, gathers, optional logistic/softmax) on representative edge accelerators.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If nearest-neighbor sampling is the main small-object bottleneck, a cheap integer bilinear surrogate or learned offset quantizer could close the remaining VisDrone APS gap without reintroducing floats.
  • The sign-dependent GELU and calibrated Shiftmax pattern is likely reusable for other deformable or multi-scale transformers (segmentation, tracking) that share the same outlier and dynamic-range problems.
  • Once the graph is fully integer, compiler auto-tuning of the per-block denominator shifts could replace the fixed mp=4 margin and further reduce QAT burden.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript presents I-LW-DETR, a fully integer-only realization of the lightweight LW-DETR detector in which every forward-pass operation — including Softmax, GELU, SiLU, LayerNorm, and deformable attention sampling — executes in integer arithmetic. The pipeline proceeds in two stages: (i) a quantization-ready FP32 variant (QR-LW-DETR) that replaces bilinear sampling in deformable attention with nearest-neighbor sampling and sinusoidal reference-box encoding with a learnable linear projection; (ii) integer approximation via three new components — a scale-preserving split convolution for the multi-scale projector, a sign-dependent ShiftGELU (SD-ShiftGELU) that avoids I-ViT's vector-wise max subtraction, and a Constrained Shiftmax with a calibrated denominator shift s_d and quotient-remainder accumulation. On COCO, QAT models lose 4.3–5.4 mAP versus FP32 LW-DETR while reducing size ~3.6× and BOPs ~13–14×; VisDrone transfer and component-wise/cumulative ablations (Tables 5–9) are provided, along with pseudocode and an operator-compatibility discussion against the Arm Ethos-U85.

Significance. If the results hold, this is a useful and timely contribution: it closes a real gap between integer-only classification/segmentation transformers (I-ViT, I-Segmenter) and the heavyweight integer-only EIQ-DETR, giving the first end-to-end integer pipeline for a lightweight deformable DETR. Strengths worth naming: the component-wise ablation (Table 5) is genuinely diagnostic — it shows naive I-ViT collapses on detectors (ShiftGELU → 0.0–8.2 mAP) and thereby justifies each proposed block; the cumulative ablation (Table 6) cleanly attributes recovery; the Constrained Shiftmax derivation (Eqs. 11–17) is internally consistent and the quotient-remainder trick is exact; full pseudocode is shipped; and the evaluation spans two external benchmarks with a standardized protocol. The hardware-mappability section is appropriately scoped (compatibility, not deployment). The work is an engineering contribution rather than a conceptual leap, but it is the kind of careful, reproducible pipeline paper the edge-deployment community uses.

major comments (3)
  1. [§3.1.1 (Integer-Compatible Deformable Attention Sampling)] The premise that bilinear interpolation is integer-incompatible is not correct as stated. Fractional weights in [0,1] quantize like any other activation: per-corner 8-bit weights with output = (Σ w_i·x_i + r) ≫ s is exactly the integer MAC pipeline used throughout the rest of the paper, and quantized resize/ROI-align are standard NPU primitives. The substitution of nearest-neighbor sampling is therefore a simplification choice, not a constraint of integer-only inference. This is load-bearing because the paper's own evidence localizes the damage where NN sampling should hurt: on VisDrone the QAT Medium model drops APS 8.8→6.1, and §5.1 concedes NN sampling 'may reduce the localization precision required for very small objects.' If an integer fixed-point bilinear recovers most of that gap, then part of the 'moderate accuracy degradation' attributed to integer arithmetic is actually the cos
  2. [§4.4 / Table 2 (QR-LW-DETR training protocol)] The training procedure for QR-LW-DETR is not specified. The learnable positional-query projector (§3.1.2) introduces new parameters, and NN sampling changes the attention function, yet §4.4 only says the 'quantization-ready architectural modifications are applied.' Were QR models fine-tuned from LW-DETR checkpoints (how many epochs, what schedule) or trained from scratch? Without this, the reported 0.6–1.7 mAP gap between LW-DETR and QR-LW-DETR (Table 2) cannot be attributed to the architectural changes themselves versus under-training of the new modules — and this gap is the baseline from which all quantization degradation is measured. Please state the QR training protocol explicitly.
  3. [§3.3.2, Eqs. (11)–(16) (robustness of calibrated s_d)] s_d is fixed per Shiftmax block from calibration statistics of S_sum, but S_sum is input-dependent at inference. If a test input produces S_sum larger than the calibrated maximum, the shifted sum S_R can exceed 2^31−1; the clamp in Eq. (16) then silently distorts the normalization factor F. Conversely the paper does not state which statistic (max? high percentile?) of S_sum is used in Eq. (11). Please (i) state the calibration statistic, (ii) report the margin between calibrated and observed S_sum on the validation set, and (iii) discuss behavior under distribution shift, since this operator is safety-critical for the whole pipeline (Table 5 shows Softmax quantization alone can take Tiny to 6.3 mAP).
minor comments (6)
  1. [Eq. (3)] 1 + 2^-1 + 2^-3 + 2^-4 = 1.6875, not 1.702. If this approximation is inherited from I-ViT, say so; otherwise the discrepancy between the stated factor and its shift implementation should be reconciled.
  2. [Table 8 / §6.4] On Tiny, SD-ShiftGELU (41.3) is only marginally above the EIQ-DETR approximation (41.0); 'consistently outperforms' overstates the Tiny case. The stronger claim is the catastrophic failure of I-ViT's variant — consider framing the comparison that way.
  3. [Tables 2–3] All results appear to be single-run; no seed variance or error bars are reported. Given that some QAT gains are ~2 mAP, at least a statement about run-to-run variability (or a second seed for one configuration) would strengthen the numbers.
  4. [§4.4] Calibration details are thin: the percentile value used for activation range estimation is not given, and the choice of 500 batches at batch size 1 is unmotivated. Also clarify which operators use the 16-bit intermediate precision in the 8/8/16 configuration and what this implies for accumulators on 8-bit-native NPUs.
  5. [§7 / Table 10] The hardware discussion is honest about being compatibility-only, but the abstract's deployment framing would benefit from one sentence noting that BOPs are a proxy and that data-dependent ops (GATHER for deformable sampling, ShiftExp) may not achieve the implied speedups; LOGISTIC/SOFTMAX primitives on Ethos-U85 being 'optional alternatives' slightly undercuts the integer-arithmetic necessity argument for SD-ShiftGELU/Shiftmax — worth a sentence.
  6. [Various] Typos/grammar: 'integrer-only' (§3 text near Figure 1), 'Contrained Shiftmax' (Table 9), 'To make the input to ShiftExp is non-positive' (§3.3.1), 'the I-LW-DETR that account for' (§7). Figure 3(b) axes are hard to read at print size. Code is promised but not yet released — please ensure the repository is public upon publication, since reproducibility of the calibration pipeline depends on it.

Circularity Check

0 steps flagged

No circularity: empirical quantization systems paper; accuracy claims are measured on external benchmarks, not forced by definition or self-citation.

full rationale

I-LW-DETR’s central claims are engineering and empirical: redesign LW-DETR for integer-only ops (nearest-neighbor deformable sampling, learnable positional queries, scale-preserving split conv, SD-ShiftGELU, Constrained Shiftmax), then report measured COCO/VisDrone mAP, size, and BOPs under PTQ/QAT against floating-point LW-DETR and prior quantized DETRs. Integer nonlinearities are explicit approximations (ShiftExp-based GELU/Softmax) whose quality is validated by ablation tables, not identities that make the reported mAP true by construction. Calibration of per-block denominator shift sd and activation ranges is standard PTQ practice and is not re-labeled as an independent prediction of the same fitted quantity. Citation of I-Segmenter (overlapping coauthors) and I-ViT appears only as related prior integer-ViT work; neither supplies a uniqueness theorem or ansatz that forces the detection results. Design choices such as NN vs bilinear are open to correctness critique but do not create a circular derivation chain. The paper is self-contained against external benchmarks; score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 4 invented entities

Load-bearing content is engineering assumptions of the I-ViT-style dyadic integer pipeline, standard symmetric quantizers, and a few hand-chosen numerical knobs (mp, Newton iters, bitwidths), plus architectural substitutes for non-integer ops. No new physical entities; invented pieces are algorithmic modules whose evidence is ablation mAP on public datasets.

free parameters (6)
  • mp (Shiftmax precision margin) = 4
    Fixed safety margin added to required log2(Ssum) when choosing denominator shift sd; applied uniformly to all blocks/sizes.
  • I-LayerNorm Newton iterations (decoder) = 20
    Increased from I-ViT’s 10 to 20 for wider decoder activation range; hand choice affecting integer variance normalization.
  • Quantization bitwidths 8/8/16 = 8/8/16
    Weights/activations 8-bit and selected nonlinear intermediates 16-bit; configuration choice that defines the reported efficiency–accuracy operating point.
  • PTQ calibration set size and range estimator = 500 batches, percentile/MA
    500 batches (bs=1), moving-average + percentile range estimation set activation scales and sd; affects all PTQ/QAT numbers.
  • QAT schedule (epochs, lr) = 15 epochs, lr=5e-7
    15 epochs at 5e-7 from calibrated PTQ with frozen observers; determines recovered mAP in main tables.
  • Common output scale for split-conv branches = max(per-branch scales)
    Chosen as max of calibrated per-branch scales so integer dyadic rescaling can sum branches.
axioms (6)
  • domain assumption Symmetric per-tensor activation and per-channel weight integer quantization with post-accumulation scales (I-ViT-style) is an adequate model of deployable integer linear ops.
    Sec. 3.2.1 adopts this as the linear backbone of the entire pipeline.
  • ad hoc to paper Nearest-neighbor sampling is an acceptable integer-compatible substitute for bilinear interpolation in deformable attention.
    Sec. 3.1.1; required for integer-only claim but changes sampling geometry, especially for small objects.
  • ad hoc to paper Learnable linear projection of reference boxes can replace sinusoidal positional encoding without breaking the two-stage query pipeline.
    Sec. 3.1.2 redesign for integer matrix-multiply compatibility.
  • domain assumption ShiftExp bit-shift exponential approximation is accurate enough on non-positive inputs when inputs are not driven too negative.
    Inherited from I-ViT; Fig. 3 and SD-ShiftGELU design assume this error structure.
  • domain assumption BOPs = sum Ops_l × ba,l × bw,l is a meaningful hardware-independent cost proxy for integer vs FP32 comparison.
    Sec. 4.2 citing Bayesian Bits; underpins ‘>10× compute’ claims without device measurements.
  • domain assumption Standard COCO/VisDrone mAP protocols and 640×640 inputs are the right success metrics for the deployment claim.
    Sec. 4.1 evaluation setup.
invented entities (4)
  • SD-ShiftGELU no independent evidence
    purpose: Integer GELU via sign-dependent sigmoid form so ShiftExp stays element-wise and non-positive without vector-max shifting.
    Sec. 3.3.1 / Alg. 1; evidence is ablation mAP vs I-ViT/I-Segmenter/EIQ GELU approximations, not external physical prediction.
  • Constrained Shiftmax (calibrated sd + quotient-remainder sum) no independent evidence
    purpose: Keep integer Softmax normalization representable in 32-bit when exponential sums are large.
    Sec. 3.3.2 / Alg. 2; validated by Softmax-only and cumulative PTQ ablations.
  • Scale-preserving split convolution (multi-branch projector) no independent evidence
    purpose: Give each multi-scale encoder feature its own activation scale before integer conv and dyadic rescaling to a common sum scale.
    Sec. 3.2.2; motivated by Fig. 7 range mismatch; gains mainly on Small/Medium.
  • QR-LW-DETR / I-LW-DETR architecture no independent evidence
    purpose: Quantization-ready then fully integer LW-DETR stack including decoder deformable path and heads.
    End-to-end system claim of the paper; performance is the empirical handle.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Enabling Fully Integer-Only Inference for Lightweight Detection Transformers." pith.science (2026). https://pith.science/paper/QAMFZALF

@misc{pith2026260724981,
  author       = {Pith},
  title        = {Pith review of: Enabling Fully Integer-Only Inference for Lightweight Detection Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QAMFZALF}},
  note         = {Machine review of arXiv:2607.24981}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Vision Transformer detectors now approach the accuracy of CNNs but remain difficult to deploy on NPUs and microcontrollers because key components, including deformable attention, feature fusion, and nonlinear activation functions, are not natively compatible with integer arithmetic. Existing quantized detectors either retain operators such as Softmax, GELU, and LayerNorm or focus on heavyweight backbones, leaving lightweight detection transformers without an end-to-end integer implementation. We address this gap with I-LW-DETR, the first fully integer-only lightweight DETR, in which every operation in the forward pass, including transformer nonlinearities, is executed in integer arithmetic. I-LW-DETR is built upon three key components: a scale-preserving split convolution that assigns independent activation scale to each branch of the multi-scale projector; SD-ShiftGELU, a sign-dependent GELU approximation that preserves element-wise behavior while avoiding the accuracy degradation; and a constrained Shiftmax that maintains stable Softmax normalization. Experimental results demonstrate that the proposed quantization pipeline consistently produces efficient fully integer-only models across different model scales. Across all model scales, the proposed pipeline incurs only a moderate accuracy degradation while reducing the model size by approximately $3.6\times$ and the computational cost by more than one order of magnitude.

Figures

Figures reproduced from arXiv: 2607.24981 by Martyna Poreba, Michal Szczepanski, Thanh Cong Le.

Figure 1
Figure 1. Figure 1: Overview of the proposed quantized I-LW-DETR pipeline. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Projector implementation before and after the proposed reformulation. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Numerical behavior of the ShiftExp approximation on negative inputs. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the ShiftGELU variants. The green block corresponds to the original [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative detection results on the COCO dataset using Small backbone. From [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative detection results on the VisDrone dataset after fine-tuning, using [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Activation distributions before and after feature concatenation. [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 1 canonical work pages

  1. [1]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, IEEE, 2016, pp. 779–788.doi:10.1109/CVPR.2016.91

    doi:10.1109/cvpr.2016.91 2016

  2. [2]

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, G. Ding, Yolov10: Real-time end-to-end object detection, in: A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, C. Zhang (Eds.), Advances in Neural Information Processing Systems, Vol. 37, Curran Associates, Inc., 2024, pp. 107984–108011.doi:10.52202/079017-3429

    doi:10.52202/079017-3429 2024

  3. [3]

    Y. Tian, Q. Ye, D. DOERMANN, Yolov12: Attention-centric real-time object detectors, in: D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Ko- niusz, M. Ghassemi, N. Chen (Eds.), Advances in Neural Information Processing Systems, Vol. 38, Curran Associates, Inc., 2025, pp. 78433– 78457. URLhttps://proceedings.neurips.cc/paper_files/paper/2025/ file/71034442590...

    arXiv 2025

  4. [4]

    Y. Zhao, Q. Zhao, C. Tu, Yolo-fb: Lightweight small-object detection via enhanced features and sparse attention, Jour- nal of Systems Architecture 176 (2026) 103777.doi:https: //doi.org/10.1016/j.sysarc.2026.103777. URLhttps://www.sciencedirect.com/science/article/pii/ S1383762126000950

    arXiv 2026

  5. [5]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Representations, 2021. URLhttps://www.bibsonomy.org/bibtex/ 2242f2231f90af37d7169530...

    2021

  6. [6]

    12346 of Lecture Notes in Computer Science, Springer, 2020, pp

    N.Carion, F.Massa, G.Synnaeve, N.Usunier, A.Kirillov, S.Zagoruyko, End-to-end object detection with transformers, in: Computer Vision – ECCV 2020, Vol. 12346 of Lecture Notes in Computer Science, Springer, 2020, pp. 213–229.doi:10.1007/978-3-030-58452-8_13. 28

    doi:10.1007/978-3-030-58452-8_13 2020

  7. [7]

    Q. Chen, X. Su, X. Zhang, J. Wang, J. Chen, Y. Shen, C. Han, Z. Chen, W. Xu, F. Li, S. Zhang, K. Yao, E. Ding, G. Zhang, J. Wang, Lw- detr: A transformer replacement to yolo for real-time detection (2024). arXiv:2406.03459. URLhttps://arxiv.org/abs/2406.03459

    Pith/arXiv arXiv 2024

  8. [8]

    Z. Yuan, C. Xue, Y. Chen, Q. Wu, G. Sun, Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization, in: Computer Vision – ECCV 2022, Vol. 13672 of Lecture Notes in Computer Science, Springer, 2022, pp. 191–207.doi:10.1007/ 978-3-031-19775-8_12

    2022

  9. [9]

    Zafrir, G

    O. Zafrir, G. Boudoukh, P. Izsak, M. Wasserblat, Q8bert: Quantized 8bit bert, in: Proceedings of the 2019 Fifth Workshop on Energy Ef- ficient Machine Learning and Cognitive Computing – NeurIPS Edition (EMC2-NIPS),IEEE,2019, pp.36–39.doi:10.1109/EMC2-NIPS53020. 2019.00016

    arXiv 2019

  10. [10]

    G. Kim, J. Lee, H. Kim, Iptq-vit: Post-training quantization of non- linear functions for integer-only vision transformers, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2026, pp. 7366–7375.doi:10.1109/WACV61042.2026.00711

    arXiv 2026

  11. [11]

    S. Kim, A. Gholami, Z. Yao, M. W. Mahoney, K. Keutzer, I-bert: Integer-only bert quantization, in: M. Meila, T. Zhang (Eds.), Proceed- ings of the 38th International Conference on Machine Learning, Vol. 139 of Proceedings of Machine Learning Research, PMLR, 2021, pp. 5506– 5518. URLhttps://proceedings.mlr.press/v139/kim21d.html

    2021

  12. [12]

    Z. Li, Q. Gu, I-ViT: Integer-only quantization for efficient vision transformer inference, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 17065–17075. URLhttps://openaccess.thecvf.com/content/ICCV2023/papers/ Li_I-ViT_Integer-only_Quantization_for_Efficient_Vision_ Transformer_Inference_ICCV_2023_paper.pdf

    2023

  13. [13]

    Sassoon, M

    J. Sassoon, M. Szczepanski, M. Poreba, I-Segmenter: Integer-only vi- sion transformer for efficient semantic segmentation, Journal of Systems Architecture 178 (2026) 103888.doi:10.1016/j.sysarc.2026.103888. 29

    arXiv 2026

  14. [14]

    S. Xu, Y. Li, M. Lin, P. Gao, G. Guo, J. Lü, B. Zhang, Q-detr: An efficient low-bit quantized detection transformer, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 3842–3851. URLhttps://openaccess.thecvf.com/content/CVPR2023/ html/Xu_Q-DETR_An_Efficient_Low-Bit_Quantized_Detection_ Transformer_CVPR_...

    2023

  15. [15]

    R. Wang, H. Sun, L. Yang, S. Lin, C. Liu, Y. Gao, Y. Hu, B. Zhang, AQ- DETR: Low-bit quantized detection transformer with auxiliary queries, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, 2024, pp. 15598–15606.doi:10.1609/aaai.v38i14.29487

    work page doi:10.1609/aaai.v38i14.29487 2024

  16. [16]

    Y. Huo, T. Wu, Y. Shen, X. Li, Z. Tao, D. Yang, QRT-DETR: Post- training quantization for real-time detection transformer, Neurocomput- ing 661 (2026) 131957.doi:10.1016/j.neucom.2025.131957

    arXiv 2026

  17. [17]

    P. Liu, C. Li, N. Zhang, J. Yang, L. Wang, Efficient integer quantization forcompressedDETRmodels, Entropy27(4)(2025)422.doi:10.3390/ e27040422

    2025

  18. [18]

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, J. Dai, Deformable detr: De- formable transformers for end-to-end object detection, in: International Conference on Learning Representations (ICLR), 2021. URLhttps://openreview.net/forum?id=gZ9hCDWe6ke

    2021

  19. [19]

    Zhang, F

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. M. Ni, H.-Y. Shum, DINO: DETR with improved denoising anchor boxes for end-to-end ob- ject detection, in: International Conference on Learning Representa- tions, 2023. URLhttps://openreview.net/forum?id=3mRwyG5one

    2023

  20. [20]

    Y. Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y. Liu, J. Chen, DETRs beat YOLOs on real-time object detection, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 16965–16974.doi:10.1109/CVPR52733.2024.01605. URLhttps://openaccess.thecvf.com/content/CVPR2024/html/ Zhao_DETRs_Beat_YOLOs_on_Real-time_Object_...

    arXiv 2024

  21. [21]

    W. Lv, Y. Zhao, Q. Chang, K. Huang, G. Wang, Y. Liu, RT-DETRv2: Improved baseline with bag-of-freebies for real-time detection trans- former (2024).arXiv:2407.17140,doi:10.48550/arXiv.2407.17140. URLhttps://arxiv.org/abs/2407.17140

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.17140 2024

  22. [22]

    Y. Peng, H. Li, P. Wu, Y. Zhang, X. Sun, F. Wu, D-FINE: Redefine regression task of DETRs as fine-grained distribution refinement, in: The Thirteenth International Conference on Learning Representations, 2025, pp. 1–18. URLhttps://openreview.net/forum?id=MFZjrTFE7h

    2025

  23. [23]

    Robinson, P

    I. Robinson, P. Robicheaux, M. Popov, D. Ramanan, N. Peri, RF- DETR: Neural architecture search for real-time detection transformers, in: International Conference on Learning Representations, 2026, pp. 1– 14

    2026

  24. [24]

    Z. Li, M. Chen, J. Xiao, Q. Gu, Psaq-vit v2: Toward accurate and gen- eral data-free quantization for vision transformers, IEEE Transactions on Neural Networks and Learning Systems 35 (12) (2024) 17227–17238. doi:10.1109/TNNLS.2023.3301007

    arXiv 2024

  25. [25]

    Y. Lin, T. Zhang, P. Sun, Z. Li, S. Zhou, Fq-vit: Post-training quan- tization for fully quantized vision transformer, in: L. D. Raedt (Ed.), Proceedings of the Thirty-First International Joint Conference on Ar- tificial Intelligence, IJCAI-22, International Joint Conferences on Ar- tificial Intelligence Organization, 2022, pp. 1173–1179, main Track. doi:...

    doi:10.24963/ijcai.2022/164 2022

  26. [26]

    Z. Li, J. Xiao, L. Yang, Q. Gu, Repq-vit: Scale reparameterization for post-training quantization of vision transformers, in: 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 17181– 17190.doi:10.1109/ICCV51070.2023.01580

    arXiv 2023

  27. [27]

    Yvinec, A

    E. Yvinec, A. Dapogny, M. Cord, K. Bailly, SPIQ: Data-free per-channel static input quantization, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 3869–3878. URLhttps://openaccess.thecvf.com/content/WACV2023/ html/Yvinec_SPIQ_Data-Free_Per-Channel_Static_Input_ Quantization_WACV_2023_paper.html 31

    2023

  28. [28]

    X. He, Y. Lu, H. Liu, C. Gong, W. He, Orq-vit: Outlier resilient post training quantization for vision transformers via outlier de- composition, Journal of Systems Architecture 168 (2025) 103530. doi:https://doi.org/10.1016/j.sysarc.2025.103530. URLhttps://www.sciencedirect.com/science/article/pii/ S1383762125002024

    arXiv 2025

  29. [29]

    T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, C. L. Zitnick, Microsoft COCO: Common objects in con- text, in: Computer Vision – ECCV 2014, Vol. 8693 of Lecture Notes in Computer Science, Springer, 2014, pp. 740–755.doi:10.1007/ 978-3-319-10602-1_48

    2014

  30. [30]

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, H. Ling, Detec- tion and tracking meet drones challenge, IEEE Transactions on Pat- tern Analysis and Machine Intelligence 44 (11) (2022) 7380–7399.doi: 10.1109/TPAMI.2021.3119563

    arXiv 2022

  31. [31]

    van Baalen, C

    M. van Baalen, C. Louizos, M. Nagel, R. A. Amjad, Y. Wang, T. Blankevoort, M. Welling, Bayesian bits: Unifying quantization and pruning, in: H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, H. Lin (Eds.), Advances in Neural Information Processing Systems, Vol. 33, Curran Associates, Inc., 2020, pp. 5741–5752. URLhttps://proceedings.neurips.cc/paper_files...

    2020

This paper was first reviewed by grok-4.5 on July 31, 2026.