Pith. sign in

REVIEW 4 major objections 6 minor 50 references

WD-DETR: Wavelet Denoising-Enhanced Real-Time Object Detection Transformer for Robot Perception with Event Cameras

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An event-only detector scores 38.4, 52.8, and 65.5 mAP on three benchmarks by denoising with wavelets inside the backbone.

desk verdict Wavelet denoising is plausible and the DSEC ablation is controlled, but the 1Mpx SOTA claim rests on a different dataset and should be re-evaluated. read the letter →

arxiv 2506.09098 v1 pith:ZKNY7TFQ submitted 2025-06-10 cs.RO cs.CV

Yangjie Cui , Boyang Gao , Yiwei Zhang , Xin Dong , Jinwu Xiang , Daochun Li , Zhan Tu This is my paper
classification cs.ROcs.CV
keywords eventcamerasobjectdetectionwaveletdenoisingDETRrepresentationreal-timeperceptionembeddeddeploymentlow-light
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

Event cameras see well in fast motion and near darkness, but the dense tensor representations used to feed neural networks accumulate noise, and the paper argues existing detectors leave that noise untreated. WD-DETR attacks the noise where features are built: each time the backbone downsamples a feature map, a wavelet transform splits it into low- and high-frequency parts, discards the high-frequency subbands, and keeps only the smoothed low-frequency approximation. With that change plus a lightweight fusion block, the event-only detector reports mean Average Precision of 38.4% on DSEC, 52.8% on Gen1, and 65.5% on 1Mpx, beating the tested event-only and event-frame fusion baselines. It also runs at roughly 35 FPS on a Jetson Orin NX, which is what makes the approach usable for onboard robot perception.

What carries the argument

The load-bearing component is wavelet pooling, built from the Haar (Daubechies) discrete wavelet transform. At each downsampling step, a 2D DWT decomposes a feature map into one low-frequency approximation $X_{ll}$ and three high-frequency detail subbands $X_{lh}$, $X_{hl}$, and $X_{hh}$; the high-frequency subbands are discarded and the inverse transform reconstructs a half-resolution map from $X_{ll}$ alone. This single operation both halves spatial resolution and filters noise, and the ablation study credits it with the largest accuracy gain. Around it, the paper wraps a time-decay event representation that converts event streams into grayscale tensors, RepConv blocks that reduce inference time, and a Dynamic Reorganization Convolution Block that lightens the hybrid encoder's cross-scale fusion.

What would settle it

Replace each wavelet-pooling downsampling step with a standard stride-2 convolution that keeps high-frequency content, and retrain on DSEC and Gen1 under identical settings; if mAP@50 does not fall by roughly the 2.6-point gap reported in the ablation, then wavelet denoising, rather than the rest of the architecture, is not what produces the gain.

Watch

Extended reading notes

Core claim

The central claim is that accumulated noise in dense event representations, rather than the sparsity of the raw event stream, is the main bottleneck for event-camera object detection, and that a wavelet-pooling downsampling layer that discards high-frequency subbands removes enough of that noise to make an event-only DETR-style detector surpass the tested state-of-the-art methods on DSEC, Gen1, and 1Mpx. Supporting this, the paper reports 38.4%, 52.8%, and 65.5% mAP respectively, with the 1Mpx result obtained on the preprocessed data from [49]; a 35 FPS TensorRT FP16 deployment on a Jetson Orin NX; and detection in low-light flight tests around 10 Lux. The authors' most direct evidence for the denoising mechanism is the DSEC ablation, where removing wavelet pooling drops mAP@50 from 38.6 to 36.0.

Load-bearing premise

The argument stands or falls on the premise that the high-frequency wavelet subbands discarded at every pooling step carry noise rather than object information, a premise tested directly only on DSEC while the 1Mpx comparison uses a differently preprocessed dataset.

Editorial extensions

If this is right

  • If the reported mAP figures hold, an event-only detector can match or exceed an events-plus-frames method on DSEC, reaching 38.4 mAP against the 38 mAP of the tested fusion approach.
  • Wavelet pooling can act as a drop-in downsampling layer that halves resolution, suppresses high-frequency noise, and adds negligible parameters, so the same design could be reused in other event-vision backbones.
  • The DRCB fusion block cuts model size by about 7 million parameters and inference time by about 8 milliseconds relative to repeated convolutions, which is what makes the 35 FPS onboard deployment possible.
  • On the 1Mpx benchmark the reported 65.5 mAP uses the preprocessed version of the dataset from [49], so the comparison is against methods evaluated on a different data preparation.
  • The outdoor flight tests at light levels near 10 Lux indicate that the combined representation and denoising pipeline can keep detecting objects where RGB cameras typically fail.

Reading between the lines

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

  • Because the denoising is embedded in a generic downsampling layer rather than a task-specific head, the same wavelet-pooling backbone could plausibly transfer to other event-camera tasks such as tracking, semantic segmentation, or depth estimation, where accumulated noise also degrades performance.
  • The event representation parameters $k$, $b$, $S_{min}$, and $S_{max}$ are not specified in the paper, so the input representation is not uniquely reproducible and reimplementations may not match the reported numbers.
  • The 1Mpx state-of-the-art comparison uses the [49]-preprocessed dataset; if competing methods were retrained on that same preprocessing, the margin could shrink, so part of the practical advantage may come from data preparation as much as from wavelet denoising.
  • A direct test of the mechanism would be to feed identical event streams to the model with and without wavelet pooling under measured noise levels; if the accuracy gap grows as noise increases, the denoising mechanism is confirmed as the source of the gain.
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 / 6 minor

Summary. The paper proposes WD-DETR, a real-time object detection transformer for event cameras. The method introduces a time-decay dense event representation, a wavelet-enhanced backbone that removes high-frequency subbands during downsampling to suppress accumulated event noise, RepConv blocks for efficiency, and a DRCB-based fusion module in the hybrid encoder. Experiments are reported on DSEC, Gen1, and 1Mpx with mAP values of 38.4%, 52.8%, and 65.5%, respectively, together with a TensorRT FP16 deployment on a Jetson Orin NX achieving about 35 FPS. The DSEC ablation (Table II) indicates that wavelet pooling improves mAP50 from 36.0 to 38.6 relative to the no-wavelet baseline, providing some evidence for the core denoising mechanism. The main caveat, acknowledged by the authors in Sec. IV-B, is that the 1Mpx evaluation uses the preprocessed dataset of [49] rather than the original benchmark used by the comparison methods.

Significance. If the reported results hold, the paper makes a useful engineering contribution to event-based object detection: it combines wavelet-based downsampling with a lightweight DETR-style detector, and the DSEC result (38.4 mAP with events only, comparable to 38 mAP of the event-frame fusion method CAFR) is encouraging. The ablation study is clearly designed and the onboard deployment demonstration on a Jetson Orin NX adds practical value. However, the headline claim of surpassing state-of-the-art methods rests principally on the 1Mpx result, which is not directly comparable to prior published numbers because of the dataset preprocessing mismatch, and the event representation is not fully specified, hindering reproducibility. The core wavelet-denoising hypothesis is validated only on DSEC, so the general claim across all three datasets remains incompletely supported. These issues are fixable but require additional work or a re-scoped claim.

major comments (4)
  1. [Sec. IV-B, Table I] The 65.5 mAP on 1Mpx and the claim of surpassing RVT-B, GET-T, and ERGO-12 are not directly comparable because WD-DETR was evaluated on the preprocessed 1Mpx dataset of [49] (which removes mosaic events and incorrect labels), whereas the comparison numbers were reported on the original 1Mpx benchmark. The manuscript itself concedes this in Sec. IV-B. The authors should either re-evaluate all comparison methods on the same preprocessed data, or present the 1Mpx result as a separate benchmark and temper the abstract and conclusion claims accordingly.
  2. [Sec. IV-A, Eqs. (2)-(5)] The event representation is not fully specified. Only the decay factor k = 1e-6 is given; the decay parameter b in Eq. (2), the clip bounds S_min and S_max in Eq. (4), the time-window/grouping length, and the exact event-to-image mapping (including how the polarity matrix P_t is accumulated and how the representation is synchronized with the annotated bounding boxes) are omitted. Without these details, the method is not uniquely reproducible. Please provide all hyperparameters or release the code.
  3. [Sec. III-C, Table II] The wavelet-denoising assumption, which discards the high-frequency subbands X_lh, X_hl, and X_hh at every wavelet pooling step, is ablated only on DSEC, where it gives a 2.6 mAP50 improvement over the no-wavelet baseline. Since the 1Mpx comparison is on a different preprocessed dataset, the DSEC ablation is currently the only direct evidence for the core denoising mechanism. Please add ablations on at least one more dataset or explicitly scope the claim to DSEC.
  4. [Table I] The inference-time comparison in Table I reports WD-DETR timing on a T4 GPU but does not specify the GPU used for the cited timings of RVT-B, S4D-ViT-B, GET-T, and ERGO-12. A cross-GPU timing comparison is not meaningful without equalizing hardware or reporting all hardware details. The separate Jetson measurement in Table V is informative and should be kept separate; the text should not imply a direct speed comparison in Table I.
minor comments (6)
  1. [Abstract and Sec. IV-B] There are typographical inconsistencies: 'WE-DETR' appears instead of 'WD-DETR' in Sec. IV-B, 'datasaet' and 'eveluation' are misspelled, and 'RTK 3090' in Sec. IV-A should likely be 'RTX 3090'.
  2. [Eq. (1)] The polarity definition states '+1 represents an increase in brightness, and +1 represents a decrease in brightness'; the second instance should be '-1'.
  3. [Table IV] The comparison between 'Our block' and 'Previous Block' should clarify that the previous block has higher GFLOPs (130.5 vs. 107.6), not just lower mAP, so the improvement is not solely due to reduced computation.
  4. [Sec. IV-D and Fig. 7] The text says 'DA IVS346' and the caption uses 'thrid view'; these should be 'DAVIS346' and 'third view', respectively.
  5. [Sec. IV-A] The training setup is incomplete: the number of epochs, batch size, learning-rate schedule, and the exact train/validation split for each dataset are not reported, which limits reproducibility.
  6. [Sec. IV-D] The outdoor flight experiment reports no quantitative detection metrics (e.g., precision, recall, or mAP) under the different lux levels; the qualitative description is acceptable as a demonstration but should be labeled as such.

Circularity Check

0 steps flagged · score 0.0 of 10

No constructive circularity: wavelet and DRCB choices are empirically ablated, and the 1Mpx caveat is a data-comparability limitation, not a derivation that reduces to its inputs.

full rationale

WD-DETR's claimed derivation chain is not circular. The event representation (Eqs. 2-5) is a defined preprocessing rule; the wavelet pooling step (Eqs. 6-10) is a fixed low-pass/downsampling operation; the DRCB and RepConv choices are architectural designs; and the detector's matching/loss (Eqs. 12-13) is a standard training objective. None of these stages defines the reported mAP into existence or fits a parameter directly to the target metric. The wavelet contribution is supported by an ablation (Table II) in which removing WaveletPooling lowers DSEC mAP50, the DRCB is supported by ablation in the same table, and the wavelet fusion configuration is compared against a previous block (Table IV); these are empirical antecedent tests, not results entailed by the method's definitions. The paper's own admission that the 1Mpx result may be partly due to using the preprocessed data of [49] ('part of the reason for this may be that the 1Mpx dataset we used was preprocessed by [49]') is an honest comparability limitation, and the underspecification of b, S_min, and S_max is a reproducibility weakness, but neither is a self-referential derivation. No load-bearing self-citation chain or uniqueness argument is invoked to forbid alternatives. The central claims rest on external benchmark measurements and ablation comparisons, so no circularity is present.

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

The central method depends on a set of unstated representation hyperparameters (decay constants and clip bounds), and the evaluation depends on the assumption that a preprocessed version of 1Mpx is comparable to the original benchmark. These are the main things a reader pays for upstream.

free parameters (4)
  • k = unspecified (decay factor reported as 1e-6)
    Decay function d(t)=(1-kt)^b in Sec. III-B controls how fast historical events fade; value not given explicitly.
  • b = unspecified
    Exponent in decay function in Eq. 2; not reported.
  • S_min, S_max = unspecified
    Clip bounds in Eq. 4 for event intensity; not reported.
  • Detection threshold = 0.5
    Used in the outdoor flight experiment only, not in benchmark evaluations.
assumptions (4)
  • domain assumption Noise is concentrated in high-frequency wavelet subbands of event representation
    Motivates discarding X_lh, X_hl, X_hh in Wavelet Pooling (Sec. III-C).
  • domain assumption Haar wavelet orthogonal filters are adequate for event feature maps
    Daubechies/Haar chosen with no comparison to other wavelets in the paper.
  • domain assumption Dense event representation can be treated as a grayscale image
    Linear mapping Eq. 5 assumes intensity can be mapped to 8-bit without losing detection-relevant information.
  • ad hoc to paper The preprocessed 1Mpx dataset from [49] is compatible with the original benchmark for comparison
    Authors use it for their method while comparing against numbers from the original dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WD-DETR: Wavelet Denoising-Enhanced Real-Time Object Detection Transformer for Robot Perception with Event Cameras." pith.science (2026). https://pith.science/paper/ZKNY7TFQ

@misc{pith2026250609098,
  author       = {Pith},
  title        = {Pith review of: WD-DETR: Wavelet Denoising-Enhanced Real-Time Object Detection Transformer for Robot Perception with Event Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKNY7TFQ}},
  note         = {Machine review of arXiv:2506.09098}
}
read the original abstract

Previous studies on event camera sensing have demonstrated certain detection performance using dense event representations. However, the accumulated noise in such dense representations has received insufficient attention, which degrades the representation quality and increases the likelihood of missed detections. To address this challenge, we propose the Wavelet Denoising-enhanced DEtection TRansformer, i.e., WD-DETR network, for event cameras. In particular, a dense event representation is presented first, which enables real-time reconstruction of events as tensors. Then, a wavelet transform method is designed to filter noise in the event representations. Such a method is integrated into the backbone for feature extraction. The extracted features are subsequently fed into a transformer-based network for object prediction. To further reduce inference time, we incorporate the Dynamic Reorganization Convolution Block (DRCB) as a fusion module within the hybrid encoder. The proposed method has been evaluated on three event-based object detection datasets, i.e., DSEC, Gen1, and 1Mpx. The results demonstrate that WD-DETR outperforms tested state-of-the-art methods. Additionally, we implement our approach on a common onboard computer for robots, the NVIDIA Jetson Orin NX, achieving a high frame rate of approximately 35 FPS using TensorRT FP16, which is exceptionally well-suited for real-time perception of onboard robotic systems.

Figures

Figures reproduced from arXiv: 2506.09098 by the authors.

Figure 1
Figure 1. Visualization of the dense event representation and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Framework of proposed WD-DETR. The event stream is split into several groups, and then be represented as a tensor [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The approach of wavelet pooling. By applying the dis [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Diverse configurations of residual modules. (a) is the basic residual block, (b) is the previous denoising block, (c) is [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The DRCB block in fusion. The DRCB block inte [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Visualization of object detection results on datasets by WD-DETR, in which event stream is densely represented again, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Outdoor flight experiment results, including the thrid view of experiment environment and the detection results of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The UAV utilized in the outdoor flight experiment is [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 37 canonical work pages

  1. [49]

    Dual memory aggregation network for event-based object detection with learnable representation,

    D. Wang, X. Jia, Y . Zhang, X. Zhang, Y . Wang, Z. Zhang, D. Wang, and H. Lu, “Dual memory aggregation network for event-based object detection with learnable representation,” inProceedings of the AAAI Conference on Artificial Intelligence, 2023

    work page 2023

  2. [1]

    Low-latency automotive vision with event cameras,

    D. Gehrig and D. Scaramuzza, “Low-latency automotive vision with event cameras,”Nature, vol. 629, no. 8014, pp. 1034–1040, 2024

    2024

  3. [2]

    E-mlb: Multilevel benchmark for event-based camera denoising,

    S. Ding, J. Chen, Y . Wang, Y . Kang, W. Song, J. Cheng, and Y . Cao, “E-mlb: Multilevel benchmark for event-based camera denoising,”IEEE Transactions on Multimedia, vol. 26, pp. 65–76, 2023

    work page 2023

  4. [3]

    A solution to the learning dilemma for recurrent networks of spiking neurons,

    G. Bellec, F. Scherr, A. Subramoney, E. Hajek, D. Salaj, R. Legenstein, and W. Maass, “A solution to the learning dilemma for recurrent networks of spiking neurons,”Nature communications, vol. 11, no. 1, p. 3625, 2020

    2020

  5. [4]

    Enabling spike-based backpropagation for training deep neural network architec- tures,

    C. Lee, S. S. Sarwar, P. Panda, G. Srinivasan, and K. Roy, “Enabling spike-based backpropagation for training deep neural network architec- tures,”Frontiers in neuroscience, vol. 14, p. 497482, 2020

    work page 2020

  6. [5]

    Graph-based spatio-temporal feature learning for neuromorphic vision sensing,

    Y . Bi, A. Chadha, A. Abbas, E. Bourtsoulatze, and Y . Andreopoulos, “Graph-based spatio-temporal feature learning for neuromorphic vision sensing,”IEEE Transactions on Image Processing, vol. 29, pp. 9084– 9098, 2020

    work page 2020

  7. [6]

    Graph-based asynchronous event processing for rapid object recog- nition,

    Y . Li, H. Zhou, B. Yang, Y . Zhang, Z. Cui, H. Bao, and G. Zhang, “Graph-based asynchronous event processing for rapid object recog- nition,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 934–943

    work page 2021

  8. [7]

    Aegnn: Asynchronous event-based graph neural networks,

    S. Schaefer, D. Gehrig, and D. Scaramuzza, “Aegnn: Asynchronous event-based graph neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 371–12 381

    2022

Show all 50 references
  1. [8]

    A voxel graph cnn for object classification with event cameras,

    Y . Deng, H. Chen, H. Liu, and Y . Li, “A voxel graph cnn for object classification with event cameras,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1172–1181

    2022

  2. [9]

    Get: Group event transformer for event-based vision,

    Y . Peng, Y . Zhang, Z. Xiong, X. Sun, and F. Wu, “Get: Group event transformer for event-based vision,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6038–6048

    2023

  3. [10]

    State space models for event cameras,

    N. Zubic, M. Gehrig, and D. Scaramuzza, “State space models for event cameras,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5819–5828

    2024

  4. [11]

    From chaos comes order: Ordering event representations for object recognition and detection,

    N. Zubi ´c, D. Gehrig, M. Gehrig, and D. Scaramuzza, “From chaos comes order: Ordering event representations for object recognition and detection,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12 846–12 856

    2023

  5. [12]

    Recurrent vision transformers for object detection with event cameras,

    M. Gehrig and D. Scaramuzza, “Recurrent vision transformers for object detection with event cameras,” inProceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, 2023, pp. 13 884– 13 893

    2023

  6. [13]

    Dsec: A stereo event camera dataset for driving scenarios,

    M. Gehrig, W. Aarents, D. Gehrig, and D. Scaramuzza, “Dsec: A stereo event camera dataset for driving scenarios,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4947–4954, 2021

    2021

  7. [14]

    A large scale event-based detection dataset for automotive,

    P. de Tournemire, D. Nitti, E. Perot, D. Migliore, and A. Sironi, “A large scale event-based detection dataset for automotive,” 2020. [Online]. Available: https://arxiv.org/abs/2001.08499

    2020 arXiv

  8. [15]

    Learning to detect objects with a 1 megapixel event camera,

    E. Perot, P. de Tournemire, D. Nitti, J. Masci, and A. Sironi, “Learning to detect objects with a 1 megapixel event camera,” 2020. [Online]. Available: https://arxiv.org/abs/2009.13436

    2020 arXiv

  9. [16]

    Fusing event-based and rgb camera for robust object detection in adverse conditions,

    A. Tomy, A. Paigwar, K. S. Mann, A. Renzaglia, and C. Laugier, “Fusing event-based and rgb camera for robust object detection in adverse conditions,” in2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 933–939

    2022

  10. [17]

    Rgb-event fusion for moving object detection in autonomous driving,

    Z. Zhou, Z. Wu, R. Boutteau, F. Yang, C. Demonceaux, and D. Ginhac, “Rgb-event fusion for moving object detection in autonomous driving,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 7808–7815

    2023

  11. [18]

    Embracing events and frames with hierarchical feature refinement network for object detection,

    H. Cao, Z. Zhang, Y . Xia, X. Li, J. Xia, G. Chen, and A. Knoll, “Embracing events and frames with hierarchical feature refinement network for object detection,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 161–177

    2025

  12. [19]

    Pushing the limits of asynchronous graph-based object detection with event cameras,

    D. Gehrig and D. Scaramuzza, “Pushing the limits of asynchronous graph-based object detection with event cameras,”arXiv preprint arXiv:2211.12324, 2022

    2022 arXiv

  13. [20]

    Memory-efficient graph con- volutional networks for object classification and detection with event cameras,

    K. Jeziorek, A. Pinna, and T. Kryjak, “Memory-efficient graph con- volutional networks for object classification and detection with event cameras,” in2023 Signal Processing: Algorithms, Architectures, Ar- rangements, and Applications (SPA). IEEE, 2023, pp. 160–165

    2023

  14. [21]

    Event-based object detection using graph neural networks,

    D. Sun and H. Ji, “Event-based object detection using graph neural networks,” in2023 IEEE 12th Data Driven Control and Learning Systems Conference (DDCLS). IEEE, 2023, pp. 1895–1900

    2023

  15. [22]

    Sodformer: Streaming object detection with transformer using events and frames,

    D. Li, Y . Tian, and J. Li, “Sodformer: Streaming object detection with transformer using events and frames,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 11, pp. 14 020–14 037, 2023

    2023

  16. [23]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,”Advances in Neural Information Processing Systems, 2017

    2017

  17. [24]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    D. Alexey, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv: 2010.11929, 2020

    2010 arXiv

  18. [25]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision. Springer, 2020, pp. 213– 229. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, DECEMBER 2024 10

    2020

  19. [26]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,”arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

  20. [27]

    Wavelet based image denoising technique,

    S. D. Ruikar and D. D. Doye, “Wavelet based image denoising technique,”International Journal of Advanced Computer Science and Applications, vol. 2, no. 3, 2011

    2011

  21. [28]

    λ-domain rate control via wavelet-based residual neural network for vvc hdr intra coding,

    F. Yuan, J. Lei, Z. Pan, B. Peng, and H. Xie, “λ-domain rate control via wavelet-based residual neural network for vvc hdr intra coding,”IEEE Transactions on Image Processing, vol. 33, pp. 6189–6203, 2024

    2024

  22. [29]

    Exploration of learned lifting- based transform structures for fully scalable and accessible wavelet-like image compression,

    X. Li, A. Naman, and D. Taubman, “Exploration of learned lifting- based transform structures for fully scalable and accessible wavelet-like image compression,”IEEE Transactions on Image Processing, vol. 33, pp. 6173–6188, 2024

    2024

  23. [31]

    Low-light image enhancement with wavelet-based diffusion models,

    H. Jiang, A. Luo, H. Fan, S. Han, and S. Liu, “Low-light image enhancement with wavelet-based diffusion models,”ACM Transactions on Graphics (TOG), vol. 42, no. 6, pp. 1–14, 2023

    2023

  24. [32]

    Research on image denoising in edge detection based on wavelet transform,

    N. You, L. Han, D. Zhu, and W. Song, “Research on image denoising in edge detection based on wavelet transform,”Applied Sciences, vol. 13, no. 3, p. 1837, 2023

    2023

  25. [33]

    Wavelet enabled convolutional autoencoder based deep neural network for hyperspectral image denoising,

    A. Paul, A. Kundu, N. Chaki, D. Dutta, and C. Jha, “Wavelet enabled convolutional autoencoder based deep neural network for hyperspectral image denoising,”Multimedia tools and applications, pp. 1–27, 2022

    2022

  26. [35]

    Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution,

    H. Huang, R. He, Z. Sun, and T. Tan, “Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution,” inProceedings of the IEEE International Conference on Computer Vision, 2017, pp. 1689–1697

    2017

  27. [36]

    Attribute enhanced face aging with wavelet-based generative adversarial networks,

    Y . Liu, Q. Li, and Z. Sun, “Attribute enhanced face aging with wavelet-based generative adversarial networks,”arXiv preprint arXiv:1809.06647, 2018

    2018 arXiv

  28. [37]

    Wavelet-enhanced convolutional neural network: A new idea in a deep learning paradigm,

    B. A. Savareh, H. Emami, M. Hajiabadi, S. M. Azimi, and M. Ghafoori, “Wavelet-enhanced convolutional neural network: A new idea in a deep learning paradigm,”Biomedical Engineering/Biomedizinische Technik, vol. 64, no. 2, pp. 195–205, 2019

    2019

  29. [38]

    Multi-level wavelet cnn for image restoration,

    P. Liu, H. Zhang, K. Zhang, L. Lin, and W. Zuo, “Multi-level wavelet cnn for image restoration,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018, pp. 773– 782

    2018

  30. [39]

    Wavelet pooling for convolutional neural networks,

    T. Williams and R. Li, “Wavelet pooling for convolutional neural networks,” inInternational Conference on Learning Representations, 2018

    2018

  31. [40]

    Sar image segmentation based on convolutional-wavelet neural network and markov random field,

    Y . Duan, F. Liu, L. Jiao, P. Zhao, and L. Zhang, “Sar image segmentation based on convolutional-wavelet neural network and markov random field,”Pattern Recognition, vol. 64, pp. 255–267, 2017

    2017

  32. [41]

    Photorealistic style transfer via wavelet transforms,

    J.-W. Yoo, Y . Uh, S. Chun, B.-K. Kang, and J.-W. Ha, “Photorealistic style transfer via wavelet transforms,”arXiv preprint arXiv:1903.09760, 2019

    1903 arXiv

  33. [42]

    Ewt: Efficient wavelet-transformer for single image denoising,

    J. Li, B. Cheng, Y . Chen, G. Gao, J. Shi, and T. Zeng, “Ewt: Efficient wavelet-transformer for single image denoising,”Neural Networks, vol. 177, p. 106378, 2024

    2024

  34. [43]

    Wavecnet: Wavelet integrated cnns to suppress aliasing effect for noise-robust image classification,

    Q. Li, L. Shen, S. Guo, and Z. Lai, “Wavecnet: Wavelet integrated cnns to suppress aliasing effect for noise-robust image classification,”IEEE Transactions on Image Processing, vol. 30, pp. 7074–7089, 2021

    2021

  35. [44]

    Repvgg: Mak- ing vgg-style convnets great again,

    X. Ding, X. Zhang, N. Ma, J. Han, G. Ding, and J. Sun, “Repvgg: Mak- ing vgg-style convnets great again,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 13 733–13 742

    2021

  36. [45]

    Impact of aliasing on generalization in deep convolutional networks,

    C. Vasconcelos, H. Larochelle, V . Dumoulin, R. Romijnders, N. Le Roux, and R. Goroshin, “Impact of aliasing on generalization in deep convolutional networks,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021, pp. 10 529–10 538

    2021

  37. [46]

    The importance of anti-aliasing in tiny object detection,

    J. Ning and M. Spratling, “The importance of anti-aliasing in tiny object detection,” inProceedings of the 15th Asian Conference on Machine Learning, ser. Proceedings of Machine Learning Research, B. Yanıko˘glu and W. Buntine, Eds., vol. 222. PMLR, 11–14 Nov 2024, pp. 975–990....

    2024

  38. [47]

    Detrs beat yolos on real-time object detection,

    Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “Detrs beat yolos on real-time object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 16 965–16 974

    2024

  39. [48]

    Lightweight Object Detection: A Study Based on YOLOv7 Integrated with ShuffleNetv2 and Vision Transformer,

    W. Gong, “Lightweight Object Detection: A Study Based on YOLOv7 Integrated with ShuffleNetv2 and Vision Transformer,” Mar. 2024

    2024

  40. [50]

    L. Sun, C. Sakaridis, J. Liang, Q. Jiang, K. Yang, P. Sun, Y . Ye, K. Wang, and L. V . Gool,Event-Based Fusion for Motion Deblurring with Cross- modal Attention. Springer Nature Switzerland, 2022, p. 412–428. [Online]. Available: http://dx.doi.org/10.1007/978-3-031-19797-0 24

    2022 doi

  41. [51]

    Rgb-event fusion for moving object detection in autonomous driving,

    Z. Zhou, Z. Wu, R. Boutteau, F. Yang, C. Demonceaux, and D. Ginhac, “Rgb-event fusion for moving object detection in autonomous driving,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 7808–7815

    2023

  42. [52]

    Mixed frame-/event-driven fast pedestrian detection,

    Z. Jiang, P. Xia, K. Huang, W. Stechele, G. Chen, Z. Bing, and A. Knoll, “Mixed frame-/event-driven fast pedestrian detection,” in2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 8332–8338

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.