Pith. sign in

REVIEW 3 major objections 5 minor 71 references

Chimera: A Block-Based Neural Architecture Search Framework for Event-Based Object Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read Chimera claims a two-stage, zero-shot neural architecture search can automatically discover compact hybrid recurrent backbones for event-based object detection, matching state-of-the-art detectors on the PEDRo dataset with an average 1.6x…

desk verdict Useful first NAS for event-based detection, but the PEDRo evaluation is compromised by selecting on the same test set used for calibration. read the letter →

arxiv 2412.19646 v1 pith:RL5MEUJX submitted 2024-12-27 cs.CV cs.AI

classification cs.CVcs.AI
keywords neuralarchitecturesearchevent-basedvisionobjectdetectionzero-shotNAShybridarchitecturesrecurrentbackbonePEDRodatasetdiversityindex
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 a neural architecture search framework built on zero-shot proxies can automatically assemble compact hybrid backbones for event-based object detection, replacing hand-designed architectures. It introduces Chimera, whose design space mixes four processing paradigms—convolutions, transformers, MLP-mixers, and state-space models—over four backbone layers plus a recurrent memory cell, and searches roughly 20,000 combinations. On the PEDRo person-detection dataset, the architectures found by the search match the accuracy of ReYOLOv8 models while using about 1.6x fewer parameters on average, and the discovered layouts transfer to the GEN1 automotive dataset without re-tuning. The intended contribution is a systematic, low-cost route from RGB-era detector designs to event-camera hardware.

What carries the argument

The load-bearing mechanism is the two-stage Chimera-NAS algorithm. Stage 1 uses an evolutionary search over a design space of four event encodings, a STEM channel count, and four Chimera layers each choosing among C2f, MaxViT, Mamba, and WaveMLP blocks, scored by a fitness function $F = \alpha W \cdot Z(f) + (1-\alpha) D(f)$ that combines Zen-Score and MACs proxies with a diversity index $D(f)$ under a parameter budget. Stage 2 trains the five highest-scoring candidates for 100 epochs and selects the one with the highest mean Average Precision. The search is carried by the zero-shot proxies: the Zen-Score, which estimates expressive capacity from a finite-difference gradient surrogate, combined 0.6/0.4 with MACs after correlation analysis on a 250-model benchmark, while the diversity index prevents the proxy bias (which favors WaveMLP) from collapsing the population into a single block type.

What would settle it

Train an equal number of randomly sampled architectures from Chimera's design space under the same 100-epoch budget on PEDRo; if the random architectures match Chimera-5M's 64.2 mAP, the zero-shot search adds nothing beyond the design space itself. Alternatively, re-running the search with the PEDRo test set withheld until after Stage 2 selection should reproduce the reported margin over ReYOLOv8s; if the margin shrinks to noise, the headline result is selection bias.

Watch

Extended reading notes

Core claim

Chimera's central claim is that a two-stage search—stage one ranking candidates with zero-shot proxies (Zen-Score, MACs, and a diversity index), stage two training the top five for 100 epochs—can find heterogeneous recurrent backbones that perform comparably to state-of-the-art event-based detectors at lower parameter counts. The strongest reported result is Chimera-5M reaching 64.2 mAP with 4.9M parameters, close to ReYOLOv8s at 64.5 mAP with 8.4M parameters, and Chimera-3M at 59.5 mAP with 3.0M parameters, slightly above ReYOLOv8n at 59.0 mAP with 4.7M parameters. The authors also show the found architectures generalize: trained from scratch on GEN1, Chimera-3M/5M/10M are competitive with similarly scaled models even though all search decisions were made on PEDRo.

Load-bearing premise

The load-bearing premise is that the PEDRo test-set mAP used to calibrate the search proxies and to select the final architecture is a genuinely held-out signal, not reuse of the same labels that guide the search.

Editorial extensions

If this is right

  • Architectures found by Chimera match hand-designed recurrent YOLO detectors on PEDRo with roughly 1.6x fewer parameters on average (59.5 mAP at 3.0M params vs 59.0 at 4.7M; 64.2 at 4.9M vs 64.5 at 8.4M).
  • The discovered layouts consistently place MaxViT early for global context, C2f for local extraction, Mamba for spatio-temporal dependencies, and WaveMLP late, suggesting a reusable ordering rule for hybrid event backbones.
  • The same searched architectures, trained from scratch on GEN1 without any search re-tuning, stay competitive with similarly sized models, showing that the PEDRo-calibrated search transfers to another event dataset.
  • The proxy calibration step identifies Zen-Score + MACs (0.6/0.4 weights) as the best cheap ranking signal on the 250-model benchmark, and NTK as too costly and negatively correlated, so future searches can drop NTK.
  • The whole first-stage search of about 1,050 models runs in roughly 1.3 hours on one GPU, plus 17.4 hours to fine-tune the top five candidates for 100 epochs each.

Reading between the lines

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

  • If the selection-bias caveat is controlled, the same two-stage recipe could be applied to other sensor domains where labeled data is scarce, such as radar or lidar detection, treating any dense encoding plus a hybrid backbone library as the search space.
  • The diversity index could serve as a general regularizer in zero-shot NAS to counter proxy bias; the paper itself observes that without it (α ≥ 0.4) the population collapses to all-WaveMLP architectures.
  • The reported 1.6x average parameter reduction is driven by the 3M and 5M models; at 10M the advantage over ReYOLOv8s essentially disappears (64.9 vs 64.5 mAP, 7.2M vs 8.4M params), so the efficiency claim is scale-dependent.
  • A stronger demonstration would be to run the search directly on GEN1 and compare the found architectures against the PEDRo-found ones, which would isolate whether the proxy ranking transfers or merely the architecture family transfers.
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

3 major / 5 minor

Summary. Chimera is a two-stage, block-based neural architecture search (NAS) framework for event-based object detection. Stage 1 searches over a design space of four macro-block types (C2f, MaxViT, Mamba, WaveMLP), four event encodings (VTEI, MDES, SHIST, TAF), and architectural hyperparameters, using a fitness function that combines zero-shot proxies (Zen-Score, MACs, and initially NTK) with a newly proposed Diversity Index. Stage 2 trains the top-five ranked architectures for 100 epochs and selects the one with the highest mean average precision (mAP). The method is evaluated on the PEDRo dataset, with 250 randomly generated heterogeneous backbone benchmarks used to calibrate the encoding, proxy weights, and diversity weight. The authors report Chimera-3M at 59.5 mAP / 3.0M parameters, Chimera-5M at 64.2 mAP / 4.9M parameters, and Chimera-10M at 64.9 mAP / 7.2M parameters, comparing favorably with ReYOLOv8n and ReYOLOv8s while claiming an average 1.6x parameter reduction. Additional results on the GEN1 dataset are provided in the Supplementary Material.

Significance. If the results hold, Chimera would be the first NAS framework specifically for event-based object detection, and it would demonstrate that zero-shot proxy search can automatically discover compact hybrid recurrent backbones that match hand-designed event detectors. The GEN1 evaluation in the Supplementary Material is a genuine strength: architectures optimized only on PEDRo are transferred to a different dataset without any adaptation and achieve competitive mAP (e.g., Chimera-3M 44.6 vs RVT-T 44.1; Chimera-10M 47.7 vs ReYOLOv8s 48.3), which supports the broad claim that the search produces transferable architectures. However, the headline PEDRo numbers are compromised by an evaluation protocol in which the same PEDRo test labels appear to be used for proxy calibration, hyperparameter selection, and final architecture selection, with no documented validation split. This selection bias could inflate the reported margins, and the GEN1 results—while promising—are less favorable than the PEDRo numbers and are currently relegated to supplementary material.

major comments (3)
  1. [§3.5, §4.1, §4.1.2, Algorithm 1] The paper does not document any validation/test split for PEDRo, yet the same PEDRo "test set" appears to be used to (i) select the SHIST encoding in §4.1 (Fig. 2), (ii) calibrate the proxy weight vector W in Table 2, (iii) tune the diversity weight α in §4.1.2, and (iv) select the final architecture in Stage 2 of Algorithm 1 ("the individual I with the highest mAP"). Consequently, the reported PEDRo mAP values in Table 12 (e.g., Chimera-5M 64.2 vs ReYOLOv8s 64.5 at 1.7x fewer parameters) are the result of selection on the test set rather than an unbiased estimate of a zero-shot NAS policy. Please introduce a proper validation split for all calibration and model-selection steps, or explicitly state the existing split and demonstrate that none of the selection steps used the test labels. The GEN1 results in Supplementary Table 13 are the only currently unbiased held-out evidence and should be moved to the main text or otherwise clearly labeled as the external validation.
  2. [Eq. (7), §3.3.1] The Diversity Index D(f) in Eq. (7) is not fully specified: S is described only as "related to the maximum value that the summation term can attain," but the claimed endpoints (D=0 for a homogeneous architecture and D=1 for a fully diverse one) hold only for a precise choice of S, such as the maximum possible pairwise-difference sum over the block-count vector for the given design space. Because D(f) is one of the two terms in the fitness function (Eq. (8)) and because α is tuned on the basis of this index, the definition of S must be made explicit for the search objective to be reproducible.
  3. [§4.1.2, §9 (Supplementary)] The calibration of α in §4.1.2 selects the value that gives the "best mAP" after training the top-five candidates for 100 epochs on PEDRo, and the claim that α=0.05 "surpasses the other configurations by approximately +2.0" is made on the same test labels that are later reported as the headline results. This is a hyperparameter choice made on the test set, not a neutral comparison of search policies. Confirming the α=0.05 choice on a validation split, or reporting the sensitivity of the final architectures to α on a held-out set, is necessary before the benefit of the diversity term can be interpreted.
minor comments (5)
  1. [§2.3] There is a typo: "ZN-NAS" should be "ZS-NAS".
  2. [§3.5] The text says "250 randomly generated heterogeneous models" and then that each model was trained with VTEI, MDES, TAF, and SHIST; the Supplementary clarifies that this yields about 1,000 trained combinations. Please state this explicitly in Section 3.5 to avoid ambiguity.
  3. [§4.1.2] The reported superiority of α=0.05 is given only as "approximately +2.0"; please report the actual mean mAP values (and, ideally, variances) for each α value.
  4. [Abstract and §4.3] The "average parameter reduction of 1.6 times" is computed against different baselines for different Chimera sizes (ReYOLOv8n for Chimera-3M, ReYOLOv8s for Chimera-5M, and so on); please state this comparison protocol explicitly rather than implying a single baseline.
  5. [Supplementary §10] Tables 7 and 8 appear to describe the same runtime data with near-identical titles; please merge them or clarify the distinction.

Circularity Check

1 steps flagged · score 4.0 of 10

PEDRo test-set mAP is used both to calibrate Chimera-NAS (encoding, proxy weights, alpha) and to select the final architecture by highest mAP, so the reported PEDRo results are selected rather than independently predicted; GEN1 provides independent support.

  1. fitted input called prediction [Section 3.5; Section 4.1; Section 4.1.2; Algorithm 1 Stage 2]
    "Each model was trained for 50 epochs, providing an effective compromise between runtime and convergence, and the mAP for the test set was recorded. ... Choose the five individuals Ik with the highest fitness scores and perform training for 100 epochs each. Output: The individual I with the highest mAP as the architecture output."

    The PEDRo test-set mAP is the common currency of every selection step. Section 3.5 records 'the mAP for the test set' for 250 benchmark models; Section 4.1 uses those values to choose SHIST and to set W=[0.6,0.4,0.0] in Table 2; Section 4.1.2 trains the top-5 architectures for each alpha and picks alpha=0.05 from their PEDRo mAP; Algorithm 1 Stage 2 then outputs 'the individual I with the highest mAP' after training the top-5. Table 12 reports these same selected models' PEDRo mAPs (59.5, 64.2, 64.9) as the headline result.

full rationale

The main circularity-adjacent issue is the reuse of the PEDRo test set for both calibration and final reporting. Section 3.5 records mAP on the PEDRo test set for 250 benchmark models; Section 4.1 uses those mAP values to select SHIST as the encoding and to set the proxy weights W; Section 4.1.2 uses the PEDRo mAP of trained top-5 architectures to choose alpha=0.05; and Algorithm 1 Stage 2 selects the final architecture as the one with the highest mAP on that same test set. The final Table 12 mAP values are therefore selected statistics rather than unbiased held-out estimates, so the headline claim of PEDRo performance comparable to ReYOLOv8s is partly driven by selection on the target metric. This is not a formal circularity in which an equation reduces to its inputs: the mAPs are empirical measurements, and the search also relies on standard external proxies (Zen-Score, MACs). The paper's own GEN1 results in the Supplementary provide genuinely independent support, because all optimization was conducted on PEDRo and the resulting architectures were evaluated from scratch on GEN1, which was not part of the search or calibration. The ReYOLOv8 backbone and baseline are self-cited, but the baseline numbers come from a separate prior paper and are not used to construct Chimera's mAP; this is incremental self-citation rather than load-bearing circularity. The paper also flags in Section 4.3 that ReYOLOv8 results exclude Random Polarity Suppression, a fairness caveat rather than a circular step. On balance, the central framework has independent content via GEN1 and the component blocks, but the PEDRo headline numbers are partially compromised by test-set reuse, giving a moderate circularity score of 4.

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

The central performance claim depends mainly on calibration choices made on PEDRo: proxy weights, alpha, and the event encoding. These are empirical fits, not derived quantities. The search also assumes the block library and recurrent backbone are an adequate design space, and that RGB-domain proxies transfer to event-domain hybrid networks. No new physical entities are introduced.

free parameters (3)
  • Proxy weight vector W = [0.6, 0.4, 0.0]
    Selected in Table 2 to maximize Kendall/Spearman correlation on the 250-model PEDRo benchmark; not derived from theory.
  • Diversity weight alpha = 0.05
    Chosen in Section 4.1.2 because it gave the highest average mAP for top-5 trained architectures on PEDRo; this is fit to the evaluation data.
  • Event encoding selection = SHIST
    SHIST was selected because it had the best average mAP and proxy correlation in the PEDRo benchmark (Figure 2); this choice affects all final results.
assumptions (4)
  • domain assumption Dense event representations preserve sufficient information for detection, and a recurrent ConvLSTM backbone is an appropriate inductive bias.
    The design space is built around ReYOLOv8's recurrent backbone and fixed YOLOv8 detection head (Section 3.1), so if dense hand-crafted event encodings or ConvLSTM are not suitable, the search cannot find a competitive detector.
  • ad hoc to paper The block library (C2f, MaxViT, Mamba, WaveMLP) is sufficient to contain high-performing event detectors.
    The search only combines these four macroblocks; no evidence is given that this library spans or approximates the space of good event backbones.
  • domain assumption Zen-Score and NTK conditioning proxies, developed for RGB CNNs, remain informative for hybrid recurrent event detectors.
    Section 3.3.1 applies Zen-Score to the whole backbone including SPPF; the benchmark itself shows limited correlation (best Kendall 0.49) and NTK has near-zero or negative correlation, so this transfer is weakly supported.
  • ad hoc to paper The Diversity Index scale factor S exists and yields D=0 for homogeneous and D=1 for fully diverse architectures.
    Equation 7 defines D only up to an unspecified scale factor S; the normalization is asserted, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chimera: A Block-Based Neural Architecture Search Framework for Event-Based Object Detection." pith.science (2026). https://pith.science/paper/RL5MEUJX

@misc{pith2026241219646,
  author       = {Pith},
  title        = {Pith review of: Chimera: A Block-Based Neural Architecture Search Framework for Event-Based Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RL5MEUJX}},
  note         = {Machine review of arXiv:2412.19646}
}
read the original abstract

Event-based cameras are sensors that simulate the human eye, offering advantages such as high-speed robustness and low power consumption. Established Deep Learning techniques have shown effectiveness in processing event data. Chimera is a Block-Based Neural Architecture Search (NAS) framework specifically designed for Event-Based Object Detection, aiming to create a systematic approach for adapting RGB-domain processing methods to the event domain. The Chimera design space is constructed from various macroblocks, including Attention blocks, Convolutions, State Space Models, and MLP-mixer-based architectures, which provide a valuable trade-off between local and global processing capabilities, as well as varying levels of complexity. The results on the PErson Detection in Robotics (PEDRo) dataset demonstrated performance levels comparable to leading state-of-the-art models, alongside an average parameter reduction of 1.6 times.

Figures

Figures reproduced from arXiv: 2412.19646 by the authors.

Figure 1
Figure 1. Structure of the Chimera Network. 3.2 Library of Components The library supporting Chimera is comprised of various building blocks and options for data encodings. This section will provide a brief overview of each component. Further details regarding their implementations are available in the Supplementary Material. 3.2.1 Building Blocks • Convolutional Layers. The well-recognized capability of Convolutional Neural … view at source ↗
Figure 2
Figure 2. a) Kendall’s Correlation between the different proxies and data formats, for the full heterogeneous architec [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Comparison between the results and the state-of-the-art for the dataset PEDRo. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: depicts the architecture of the C2f block [48], an optimized variant of the Cross-Stage Partial (CSP) Bottleneck block featuring two convolutional layers. The first convolutional layer adapts the input channel count. Following this, a Split block divides the features i…
Figure 5
Figure 5. Figure 5: MaxViT block structure [16]. 5.3 Mamba Block Mamba is a block based on State Space Models (SSM). A continuous-time SSM can be defined according to the following equations: h ′ (t) = Ah(t) + Bx(t) (9) y(t) = Ch(t) (10) In these equations, h(t) ∈ RM, x(t) ∈ R, and y(t) ∈…
Figure 6
Figure 6. Figure 6: Mamba block. [20] In the original implementation, the Mamba block alternates between a State Space Model (SSM) and Self-Attention mechanisms within the same stage. However, in this paper, we have decided to retain only the SSM block, as MaxViT already incorporates Self…
Figure 7
Figure 7. Figure 7: WaveMLP main block [54]. Up Concat Up C2f_1 Concat C2f_2 Concat Conv 2D_1 C2f_3 Concat Conv 2D_2 C2f_4 P5 P4 P3 Detec�on Head 1 Detec�on Head 2 Detec�on Head 3 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Multi-scale feature fusion and Detection Head structures from YOLOv8. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Mixed-Density Event Stacks (MDES). 5 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Volume of Ternary Event Image (VTEI). 7 More details about training The training hyperparameters and procedures implemented in this study were primarily adapted from ReYOLOv8 [10], and YOLOv8 [48], with minor modifications to batch sizes and learning rates [PITH_FULL…
Figure 11
Figure 11. Figure 11: mAP distribution for the different model compositions. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Distribution of the Zen Score for the different categories of models. [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Average mAP of the Top-5 performers and Average Diversity of populations after evolving over [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Comparison between the results and the state-of-the-art for the dataset GEN1. [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 51 canonical work pages

  1. [1]

    Benchmarking robustness in object detection: Autonomous driving when winter is coming

    Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019

  2. [2]

    The object detection, perspective and obstacles in robotic: a review

    Ge Xu, Amir Sohail Khan, Ata Jahangir Moshayedi, Xiaohong Zhang, and Yang Shuxin. The object detection, perspective and obstacles in robotic: a review. EAI Endorsed Transactions on AI and Robotics, 1(1), 2022

  3. [3]

    Real time object detection and trackingsys- tem for video surveillance system

    Sudan Jha, Changho Seo, Eunmok Yang, and Gyanendra Prasad Joshi. Real time object detection and trackingsys- tem for video surveillance system. Multimedia Tools and Applications, 80(3):3981–3996, 2021

  4. [4]

    Deep learning for generic object detection: A survey

    Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikäinen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261–318, 2020. 10 A PREPRINT - D ECEMBER 30, 2024

  5. [5]

    A 128 × 128 120 db 15 µs latency asynchronous temporal contrast vision sensor

    Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A 128 × 128 120 db 15 µs latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43(2):566–576, 2008

  6. [6]

    Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza

    Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154–180, 2022

  7. [7]

    A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas

    Juan Terven, Diana-Margarita Córdova-Esparza, and Julio-Alejandro Romero-González. A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas. Machine Learning and Knowledge Extraction, 5(4):1680–1716, 2023

  8. [8]

    Learning to detect objects with a 1 megapixel event camera

    Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. Advances in Neural Information Processing Systems, 33:16639–16652, 2020

Show all 71 references
  1. [9]

    Asynchronous spatio-temporal memory network for continuous event-based object detection

    Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975– 2987, 2022

  2. [10]

    A recurrent yolov8-based framework for event-based object detection

    Diego A Silva, Kamilya Smagulova, Ahmed Elsheikh, Mohammed E Fouda, and Ahmed M Eltawil. A recurrent yolov8-based framework for event-based object detection. arXiv preprint arXiv:2408.05321, 2024

  3. [11]

    Recurrent vision transformers for object detection with event cameras

    Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13884–13893, 2023

  4. [12]

    Scene adaptive sparse transformer for event-based object detection

    Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794–16804, 2024

  5. [13]

    Get: group event transformer for event- based vision

    Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: group event transformer for event- based vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6038–6048, 2023

  6. [14]

    State space models for event cameras

    Nikola Zubic, Mathias Gehrig, and Davide Scaramuzza. State space models for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5819–5828, 2024

  7. [15]

    Fastervit: Fast vision transformers with hierarchical attention

    Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. arXiv preprint arXiv:2306.06189, 2023

  8. [16]

    Maxvit: Multi-axis vision transformer

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European conference on computer vision, pages 459–479. Springer, 2022

  9. [17]

    Edgevit: Efficient visual modeling for edge computing

    Zekai Chen, Fangtian Zhong, Qi Luo, Xiao Zhang, and Yanwei Zheng. Edgevit: Efficient visual modeling for edge computing. In International Conference on Wireless Algorithms, Systems, and Applications, pages 393–405. Springer, 2022

  10. [18]

    Efficientformer: Vision transformers at mobilenet speed

    Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35:12934–12949, 2022

  11. [19]

    Convmlp: Hierarchical convolutional mlps for vision

    Jiachen Li, Ali Hassani, Steven Walton, and Humphrey Shi. Convmlp: Hierarchical convolutional mlps for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6307–6316, 2023

  12. [20]

    Mambavision: A hybrid mamba-transformer vision backbone

    Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024

  13. [21]

    A comprehensive survey of neural architecture search: Challenges and solutions.ACM Computing Surveys (CSUR), 54(4):1–34, 2021

    Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Xiaojiang Chen, and Xin Wang. A comprehensive survey of neural architecture search: Challenges and solutions.ACM Computing Surveys (CSUR), 54(4):1–34, 2021

  14. [22]

    Carbon emissions and large neural network training.arXiv preprint arXiv:2104.10350, 2021

    David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. Carbon emissions and large neural network training.arXiv preprint arXiv:2104.10350, 2021

  15. [23]

    Aegnn: Asynchronous event-based graph neural networks

    Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12371– 12381, 2022

  16. [24]

    Event-based object detection using graph neural networks

    Daobo Sun and Haibo Ji. Event-based object detection using graph neural networks. In 2023 IEEE 12th Data Driven Control and Learning Systems Conference (DDCLS), pages 1895–1900. IEEE, 2023. 11 A PREPRINT - D ECEMBER 30, 2024

  17. [25]

    Low-latency automotive vision with event cameras

    Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034–1040, 2024

  18. [26]

    Hybrid snn-ann: Energy-efficient classification and object detection for event-based vision

    Alexander Kugele, Thomas Pfeil, Michael Pfeiffer, and Elisabetta Chicca. Hybrid snn-ann: Energy-efficient classification and object detection for event-based vision. In DAGM German Conference on Pattern Recognition, pages 297–312. Springer, 2021

  19. [27]

    Object detection with spiking neural networks on automotive event data

    Loïc Cordone, Benoît Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2022

  20. [28]

    Deep directly-trained spiking neural networks for object detection

    Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555–6565, 2023

  21. [29]

    Direct training high-performance spiking neural networks for object recognition and detection

    Yu Zhang, Hong Zhang, Yang Li, Bin He, Xiongfei Fan, and Yue Wang. Direct training high-performance spiking neural networks for object recognition and detection. Frontiers in Neuroscience, 17:1229951, 2023

  22. [30]

    Spike-event object detection for neuromorphic vision.IEEE Access, 11:5215–5230, 2023

    Yuan-Kai Wang, Shao-En Wang, and Ping-Hsien Wu. Spike-event object detection for neuromorphic vision.IEEE Access, 11:5215–5230, 2023

  23. [31]

    Object detection for embedded systems using tiny spiking neural networks: Filtering noise through visual attention

    Hugo Bulzomi, Amélie Gruel, Jean Martinet, Takeshi Fujita, Yuta Nakano, and Rémy Bendahan. Object detection for embedded systems using tiny spiking neural networks: Filtering noise through visual attention. In 2023 18th International Conference on Machine Vision and Applicatio...

  24. [32]

    Sfod: Spiking fusion object detector

    Yimeng Fan, Wei Zhang, Changsong Liu, Mingyang Li, and Wenrui Lu. Sfod: Spiking fusion object detector. arXiv preprint arXiv:2403.15192, 2024

  25. [33]

    Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks

    Ziming Wang, Ziling Wang, Huaning Li, Lang Qin, Runhao Jiang, De Ma, and Huajin Tang. Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks. arXiv preprint arXiv:2403.12574, 2024

  26. [34]

    Better and faster: Adaptive event conversion for event-based object detection

    Yansong Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2056–2064, 2023

  27. [36]

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

    Nikola Zubi´c, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12846–12856, 2023

  28. [37]

    Hierarchical neural memory network for low latency event processing

    Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22867–22876, 2023

  29. [38]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023

  30. [39]

    Attention is all you need

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

  31. [40]

    Escaping the big data paradigm with compact transformers

    Ali Hassani, Steven Walton, Nikhil Shah, Abulikemu Abuduweili, Jiachen Li, and Humphrey Shi. Escaping the big data paradigm with compact transformers. arXiv preprint arXiv:2104.05704, 2021

  32. [41]

    Efficientvmamba: Atrous selective scan for light weight visual mamba

    Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024

  33. [42]

    Bottleneck transformers for visual recognition

    Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16519–16529, 2021

  34. [43]

    Crossvit: Cross-attention multi-scale vision trans- former for image classification

    Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision trans- former for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021

  35. [44]

    Co-scale conv-attentional image transformers

    Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Co-scale conv-attentional image transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9981–9990, 2021

  36. [45]

    Visformer: The vision-friendly transformer

    Zhengsu Chen, Lingxi Xie, Jianwei Niu, Xuefeng Liu, Longhui Wei, and Qi Tian. Visformer: The vision-friendly transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 589–598, 2021. 12 A PREPRINT - D ECEMBER 30, 2024

  37. [46]

    Cvt: Introducing convolutions to vision transformers

    Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22–31, 2021

  38. [47]

    Zen-nas: A zero-shot nas for high-performance image recognition

    Ming Lin, Pichao Wang, Zhenhong Sun, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, and Rong Jin. Zen-nas: A zero-shot nas for high-performance image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 347–356, 2021

  39. [48]

    Ultralytics YOLO, January 2023

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLO, January 2023

  40. [49]

    Convolutional lstm network: A machine learning approach for precipitation nowcasting

    Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional lstm network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015

  41. [50]

    Learning to forget: Continual prediction with lstm.Neural computation, 12(10):2451–2471, 2000

    Felix A Gers, Jürgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with lstm.Neural computation, 12(10):2451–2471, 2000

  42. [51]

    Spatial pyramid pooling in deep convolutional networks for visual recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904–1916, 2015

  43. [52]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012

  44. [53]

    Mlp-mixer: An all-mlp architecture for vision

    Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in neural information processing systems, 34:242...

  45. [54]

    An image patch is a wave: Phase-aware vision mlp

    Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Phase-aware vision mlp. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10935–10944, 2022

  46. [55]

    Stereo depth from events cameras: Concentrate and focus on the future

    Yeongwoo Nam, Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Stereo depth from events cameras: Concentrate and focus on the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6114–6123, 2022

  47. [56]

    Motion robust high-speed light-weighted object detection with event camera

    Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 2023

  48. [57]

    Zero-shot neural architecture search: Challenges, solutions, and opportunities

    Guihong Li, Duc Hoang, Kartikeya Bhardwaj, Ming Lin, Zhangyang Wang, and Radu Marculescu. Zero-shot neural architecture search: Challenges, solutions, and opportunities. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

  49. [58]

    Neural architecture search on imagenet in four gpu hours: A theoretically inspired perspective

    Wuyang Chen, Xinyu Gong, and Zhangyang Wang. Neural architecture search on imagenet in four gpu hours: A theoretically inspired perspective. arXiv preprint arXiv:2102.11535, 2021

  50. [59]

    Pedro: an event-based dataset for person detection in robotics

    Chiara Boretti, Philippe Bich, Fabio Pareschi, Luciano Prono, Riccardo Rovatti, and Gianluca Setti. Pedro: an event-based dataset for person detection in robotics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064–4069, 2023

  51. [60]

    A large scale event-based detection dataset for automotive

    Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020

  52. [61]

    Backpropagation through time and derivative adaptive critics: A common framework for comparison

    Danil V Prokhorov. Backpropagation through time and derivative adaptive critics: A common framework for comparison. Handbook of learning and approximate dynamic programming, 45:381–404, 2004. 13 A PREPRINT - D ECEMBER 30, 2024 Chimera: A Block-Based Neural Architecture Search ...

  53. [62]

    CNN + RNN 4.7M 59.0 Chimera-5M (this work) Hybrid + RNN 4.9M 64.2 Chimera-10M (this work) Hybrid + RNN 7.2M 64.9 WaveMLP-Baseline (Based on [20]) MLP-Mixer + RNN 8.2M 59.7 ReYOLOv8s

  54. [63]

    CNN + RNN 8.4M 64.5 MaxViT-Baseline (Based on [11]) Transformer + RNN 9.2M 52.6 Mamba-Baseline (Based on [20]) Mamba + RNN 14.0M 46.9 ReYOLOv8m

  55. [64]

    CNN + RNN 18.1M 66.5 YOLOv8x

  56. [65]

    CNN 68.2M 58.6 Figure 14: Comparison between the results and the state-of-the-art for the dataset GEN1. Since the primary focus of this paper of establishing procedures for creating hybrid architectures was fully carried out through PEDRo, the results for GEN1 were reserved fo...

  57. [66]

    Transformer + RNN 4.4M 44.1 ReYOLOv8n

  58. [67]

    CNN + RNN 4.7M 46.3 Chimera-5M (this work) Hybrid + RNN 4.9M 46.0 Chimera-10M (this work) Hybrid + RNN 7.2M 47.7 ReYOLOv8s

  59. [68]

    CNN + RNN 8.4M 48.3 HMNet-B1

  60. [69]

    HMNet 9.4M 45.5 S5-ViT-S

  61. [70]

    Transformer + SSM 9.7M 46.6 RVT-S

  62. [71]

    Transformer + RNN 9.9M 46.5 HMNet-L1

  63. [72]

    HMNet 11.4M 47.0 AED

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.