From Ground to Air: Noise Robustness in Vision Transformers and CNNs for Event-Based Vehicle Classification with Potential UAV Applications

Hector Figueiredo; Miguel Arana-Catania; Nouf Almesafri

REVIEW 4 major objections 4 minor 16 references

From Ground to Air: Noise Robustness in Vision Transformers and CNNs for Event-Based Vehicle Classification with Potential UAV Applications

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that on event-based vehicle classification, Vision Transformer B16 beats ResNet34 under spatial shift, event loss, and polarity-reversal noise, even though ResNet34 is slightly better on clean data.

desk verdict The headline robustness claim is not supported because ResNet34 and ViT B16 are evaluated on different test sets and with different training budgets, so the paper needs major revision before the result can be trusted. read the letter →

arxiv 2506.22360 v1 pith:Q3CKSJM2 submitted 2025-06-27 cs.CV cs.AIcs.LG

Nouf Almesafri , Hector Figueiredo , Miguel Arana-Catania This is my paper

classification cs.CVcs.AIcs.LG

keywords event-basedvisiontransformerResNet34noiserobustnessvehicleclassificationGEN1datasetUAVapplicationsEventSpikeTensor

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

This paper asks which architecture—CNN or transformer—handles the noisy, asynchronous output of event-based cameras better for vehicle classification. Fine-tuning ResNet34 and ViT B16 on the GEN1 dataset, the authors report that ResNet34 is more accurate on clean data (88% vs 86%), while ViT B16 degrades more gracefully under all three simulated noise types: spatial shifts, event loss, and polarity reversal. If true, the result means transformer-based models can be the more reliable choice when event streams are imperfect, which is exactly the condition UAVs and autonomous vehicles face. The practical upshot is that a small clean-data accuracy advantage should not by itself decide the architecture choice for noisy, real-world perception.

What carries the argument

The central machinery is the Event Spike Tensor (EST) representation, which packs asynchronous events into a grid tensor with spatial, temporal, and polarity dimensions, together with three noise-injection formulas: uniform spatial shifts with $\Delta x, \Delta y \sim U(-\delta, \delta)$, event dropping at rate $\eta$, and polarity flipping at probability $\rho$. ResNet34 and ViT B16 are fine-tuned on this tensor input, and robustness is measured by accuracy, ROC AUC, and average precision across noise levels. The argument is carried by the accuracy-versus-noise curves, where ViT B16's slower decline is attributed to transformer self-attention not being anchored to exact pixel locations.

What would settle it

Evaluate both models on identical random subsets of the GEN1 test set (e.g., 3672 samples each), using the same number of epochs and the same noise realizations; if ViT B16 no longer retains its accuracy lead at 20% shift, loss, or polarity-reversal noise, the claimed robustness advantage is not architecture-driven.

Watch

Extended reading notes

Core claim

The central discovery reported is that ViT B16 surpasses ResNet34 in noisy environments, showing greater robustness to all noise types. On clean GEN1 data, ResNet34 reaches 88% accuracy and ViT B16 reaches 86%, but as event shift, event loss, and polarity reversal intensify, ViT B16's accuracy declines more slowly, with the largest gap appearing under x/y spatial shifts. The authors interpret this as ViT B16 generalizing better in real-world noisy applications because its patch-based self-attention depends less on exact spatial locations than convolutional filters do. They position the finding as making ViT B16 the more reliable choice for dynamic environments such as UAVs, with the trade-off of slightly lower accuracy on clean frames.

Load-bearing premise

The load-bearing premise is that ResNet34 and ViT B16 were compared under equivalent evaluation conditions; the reported test supports (3672 vs 7720) and epoch counts (35 vs 10) differ, so the observed robustness advantage may be an artifact of these discrepancies rather than of architecture.

Editorial extensions

If this is right

If ViT B16's robustness holds, event-based perception systems for UAVs and autonomous vehicles should prefer transformer backbones whenever noise is expected, accepting a small clean-accuracy trade-off.
Spatial-shift noise is the clearest differentiator, so platforms with imperfect sensor calibration or motion artifacts stand to gain most from ViT B16.
The robustness appears without event-specific pretraining, suggesting that general-purpose transformer features transfer to event-based vehicle classification.
Event loss and polarity reversal hurt both models, so sensor-level preprocessing that reduces these errors should help regardless of architecture.

Reading between the lines

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

A reader's extension: because the reported test supports differ (3672 samples for ResNet34 versus 7720 for ViT B16), the robustness margin should be re-checked on equal-size test sets before drawing deployment conclusions.
The same sensor-level noise pipeline could be applied to aerial classes such as drones, birds, and aircraft; the noise model does not depend on object class, so the robustness ranking likely transfers.
Comparing transformer variants with different patch sizes would test whether the noise resilience scales with patch size or model capacity.
Since ViT B16 uses only ImageNet pretraining, the results hint that dedicated event-based pretraining is not a prerequisite for noise robustness, lowering the barrier for event-based deployments.

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.

Desk Editor's Note

read the letter

The specific comparison you ask about — ResNet34 versus ViT B16 on GEN1 under spatial shift, event loss, and polarity reversal — is a reasonable empirical question, and the paper deserves credit for laying out a clean noise taxonomy and for reporting clean accuracy, confusion matrices, ROC, and PR curves. The clean numbers (88% vs 86%) are at least plausible. But the central claim in Section III.D, that ViT B16 is more robust to all noise types, is not supported by the paper's own evidence. The stress-test note is on target: ResNet34's classification report shows a test support of 3672 (Table 3), while ViT B16's shows 7720 (Table 7). Both models supposedly use the same 70:15:15 split, so these cannot both be the held-out test set. The accuracy curves in Figure 17 may therefore be comparing different test distributions, which alone would invalidate the head-to-head conclusion. It is also not a minor inconsistency: it is load-bearing because the paper's only real result is the crossover in robustness. Training conditions also differ: ResNet34 ran 35 epochs across 5 folds with best-fold selection; ViT B16 ran 10 epochs with dropout selection. Section II.F says 50 epochs, batch size 8, and learning rate 1e-4, but the tables list 35/10 epochs, batch size 10 for ViT, and learning rate 2.2e-5. That is a tangle of mismatches, and any one of them could explain the observed differences. The noise evaluation protocol is also ambiguous: Section II.F says noise was added to the training and validation sets, not the test set. If that is literal, the models were trained on corrupted data, so the robustness result may reflect augmentation rather than architecture. Finally, the quantitative noise results are missing: Table 8 is empty, Table 9 is empty, and the only evidence is a figure without error bars or confidence intervals. That is a serious omission for a benchmark paper. On the positive side, the paper is clearly written and the clean-data evaluation is mostly transparent. The authors also acknowledge their own limitations regarding ViT pretraining and dataset size. But the manuscript as it stands is not coherent on its own terms: the reported numbers contradict each other, and the empty tables remove any possibility of independent verification. This is the kind of paper that should be desk-rejected or sent back for major revision before reviewers spend time on it. If the authors fix the evaluation to use identical test splits and training budgets, fill in the tables, clarify the noise protocol, and release code and data, the question would be worth revisiting. As submitted, I would not cite it and would not bring it to a reading group, because the central conclusion is not established.

Referee Report

4 major / 4 minor

Summary. The paper compares a ResNet34 CNN and a ViT B16 transformer for vehicle/pedestrian classification on event-based data from the GEN1 dataset, under clean conditions and under three simulated noise types (spatial shift, event loss, polarity reversal). Clean accuracy is reported as 88% for ResNet34 and 86% for ViT B16; the paper's headline claim in Section III.D is that ViT B16 surpasses ResNet34 in all noisy conditions despite slightly lower clean accuracy, with implications for UAV applications.

Significance. If the headline claim were established, it would be a useful data point for architecture selection in event-based vision, particularly for deployment environments with degraded event streams. The paper selects a relevant benchmark (GEN1), defines three practically motivated noise models, and reports both ROC/PR metrics. However, the comparison is not controlled: the two models are evaluated on test sets of different sizes, trained for different numbers of epochs with different batch sizes and model-selection protocols, and the quantitative noise results are not reported in tabular form. These issues directly undermine the central robustness claim, so the paper in its current form does not provide a reliable empirical finding.

major comments (4)

[§II.F vs Tables 2 and 5] The training protocol stated in §II.F (50 epochs, learning rate 1e-4, batch size 8) contradicts the values reported in Table 2 (35 epochs, learning rate 2.2e-5, batch size 8) and Table 5 (10 epochs, learning rate 2.2e-5, batch size 10). Since the paper claims a fair comparison, this inconsistency in the experimental setup must be resolved before any architecture comparison can be trusted.
[Tables 3 and 7] The test supports for the clean evaluation are 3672 for ResNet34 (Table 3) and 7720 for ViT B16 (Table 7). Section II.F states both models use a 70:15:15 split of the same GEN1 dataset; identical splits should produce identical test-set sizes. The factor-of-two difference indicates that the test sets are not the same, so the clean accuracy values (88% vs 86%) and the noise curves in Figure 17 are not directly comparable. This invalidates the load-bearing conclusion in Section III.D that ViT B16 is more robust to all noise types.
[§III.C and Table 8] The only quantitative noise evidence is Figure 17; Table 8, which is supposed to summarize performance at different noise levels, is empty. Without per-level accuracy values, test supports, or confidence intervals, the reader cannot verify the claimed crossover. Furthermore, Section II.F states that 'noise was systematically added to the training and validation sets,' which means the models may have been trained on noisy data. If so, the measured degradation is a blend of training-time and test-time noise effects, and the robustness comparison does not isolate architectural behavior.
[§III.A and §III.B] The model selection procedures differ: ResNet34 uses 5-fold cross-validation and selects the best fold, while ViT B16 selects between dropout configurations on the validation set. This is not a matched protocol; either model could be favored by its selection procedure. If the intended claim is architectural, the evaluation should use the same split, the same selection criterion, and ideally the same number of training epochs, or the discrepancy should be explicitly analyzed.

minor comments (4)

[Table 8 caption] The caption mentions 'N-Caltech101, GEN1, and noisy GEN1 datasets,' but the manuscript never reports any experiments on N-Caltech101; either include those results or remove the reference.
[Abstract and §III.B] The text repeatedly says ViT B16 was 'pre-trained on a smaller dataset' compared to ResNet34, but both models are initialized with ImageNet weights; the difference in pre-training data size is not explained and appears to be a misunderstanding of the ViT literature.
[Figure 17] Figure 17 is never called out with specific panel references in the text; please indicate which subplots correspond to shift, event loss, and polarity reversal, and add axis labels and error bars if available.
[§II.F] The stated learning rate and training epochs in the text differ from the tables as noted above, but even internally the sentence about '50 epochs with early stopping' is not consistent with the reported early stopping in Tables 2 and 5.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical benchmark whose robustness comparison rests on measured accuracies, not on fitted inputs or self-citations.

full rationale

The paper is an empirical benchmark comparing ResNet34 and ViT B16 on clean and noisy GEN1 event-based data. No derivation chain is claimed: there is no theoretical model, no fitted parameter that is later renamed as a prediction, and no result that is defined in terms of the quantity it is supposed to establish. The architecture choices and the EST representation are cited from external prior work (He et al. 2016; Dosovitskiy et al. 2020; Gehrig et al. 2019), and none of those citations is authored by the present paper's authors, so no self-citation chain is load-bearing. The central robustness claim in Section III.D, that ViT B16 surpasses ResNet34 in noisy environments, is presented as a direct reading of measured accuracy curves (Figure 17), not as a consequence of any equation or fitted input. The reported clean accuracies, supports, and training details are empirical observations, not constructions that presuppose the conclusion. There is a legitimate experimental-validity concern that the two models appear to be evaluated on test sets of different sizes (Table 3 support 3672 vs Table 7 support 7720) despite the stated 70:15:15 split, and the training budgets differ (35 epochs vs 10 epochs). However, that is a fairness or reproducibility issue, not circularity: the accuracy numbers are not shown to be equivalent to their inputs by construction. No step in the paper reduces its conclusion to a fit, a self-citation, or a definitional identity, so the appropriate circularity score is 0.

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

The paper contributes no new theoretical objects. It relies on standard pretrained architectures, a standard event representation, simulated noise models, and the GEN1 benchmark. The main added assumptions are the transferability of ImageNet features and the adequacy of the simulated noise protocol.

free parameters (4)

Learning rate = 2.2e-5
Selected via Optuna on validation data; used for both models.
Training epochs = ResNet34: 35, ViT B16: 10
The models receive very different training budgets, which can explain accuracy differences; stated in Tables 2 and 5, conflicting with Section II.F's 50 epochs.
Batch size = ResNet34: 8, ViT B16: 10
Different batch sizes across models; another training condition difference.
Dropout layers (ViT) = 2
Selected based on validation accuracy; affects the ViT result.

assumptions (4)

domain assumption ImageNet-pretrained weights generalize to event-based representations
Both models are initialized with ImageNet weights (Section II.F); the study assumes this transfer is valid and equal for both architectures.
domain assumption Simulated noise (spatial shift, event loss, polarity reversal) approximates real event-camera noise
Section II.E introduces these corruptions as representative real-world challenges, but no validation against real noisy event data is provided.
domain assumption EST representation preserves enough information for the classification task
Section II.C relies on EST; no ablation on representation choice is given.
domain assumption GEN1 dataset splits are representative and correctly used
Different test set sizes for the two models suggest this assumption is violated; see red flags.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Ground to Air: Noise Robustness in Vision Transformers and CNNs for Event-Based Vehicle Classification with Potential UAV Applications." pith.science (2026). https://pith.science/paper/Q3CKSJM2

@misc{pith2026250622360,
  author       = {Pith},
  title        = {Pith review of: From Ground to Air: Noise Robustness in Vision Transformers and CNNs for Event-Based Vehicle Classification with Potential UAV Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q3CKSJM2}},
  note         = {Machine review of arXiv:2506.22360}
}

read the original abstract

This study investigates the performance of the two most relevant computer vision deep learning architectures, Convolutional Neural Network and Vision Transformer, for event-based cameras. These cameras capture scene changes, unlike traditional frame-based cameras with capture static images, and are particularly suited for dynamic environments such as UAVs and autonomous vehicles. The deep learning models studied in this work are ResNet34 and ViT B16, fine-tuned on the GEN1 event-based dataset. The research evaluates and compares these models under both standard conditions and in the presence of simulated noise. Initial evaluations on the clean GEN1 dataset reveal that ResNet34 and ViT B16 achieve accuracies of 88% and 86%, respectively, with ResNet34 showing a slight advantage in classification accuracy. However, the ViT B16 model demonstrates notable robustness, particularly given its pre-training on a smaller dataset. Although this study focuses on ground-based vehicle classification, the methodologies and findings hold significant promise for adaptation to UAV contexts, including aerial object classification and event-based vision systems for aviation-related tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages

[1]

Retinomorphic event-based vision sensors: bioinspired cameras with spiking output,

Posch, C., Serrano-Gotarredona, T., Linares-Barranco, B., and Delbruck, T., “Retinomorphic event-based vision sensors: bioinspired cameras with spiking output,”Proceedings of the IEEE, Vol. 102, No. 10, 2014, pp. 1470–1484

work page 2014
[2]

Collision detection for UAVs using event cameras,

Paulo, J. P. C., “Collision detection for UAVs using event cameras,” Ph.D. thesis, University of Lisbon, 2023

work page 2023
[3]

Computer vision for autonomous vehicles: Problems, datasets and state of the art,

Janai, J., Güney, F., Behl, A., Geiger, A., et al., “Computer vision for autonomous vehicles: Problems, datasets and state of the art,”Foundations and Trends®in Computer Graphics and Vision, Vol. 12, No. 1–3, 2020, pp. 1–308

work page 2020
[4]

Deep residual learning for image recognition,

He, K., Zhang, X., Ren, S., and Sun, J., “Deep residual learning for image recognition,”Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

work page 2016
[5]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv preprint arXiv:2010.11929, 2020

arXiv 2010
[6]

A large scale event-based detection dataset for automotive,

De Tournemire, P., Nitti, D., Perot, E., Migliore, D., and Sironi, A., “A large scale event-based detection dataset for automotive,” arXiv preprint arXiv:2001.08499, 2020

arXiv 2001
[7]

Discussion on event-based cameras for dynamic obstacles recognition and detection for UAVs in outdoor environments,

Ben Miled, M., Zeng, Q., and Liu, Y., “Discussion on event-based cameras for dynamic obstacles recognition and detection for UAVs in outdoor environments,”UKRAS22 Conference" Robotics for Unconstrained Environments" Proceedings, EPSRC UK-RAS Network, 2022, pp. 34–35

work page 2022
[8]

End-to-end learning of representations for asynchronous event-based data,

Gehrig, D., Loquercio, A., Derpanis, K. G., and Scaramuzza, D., “End-to-end learning of representations for asynchronous event-based data,”Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 5633–5643

work page 2019

Show all 16 references

[9]

A differentiable recurrent surface for asynchronous event-based data,

Cannici, M., Ciccone, M., Romanoni, A., and Matteucci, M., “A differentiable recurrent surface for asynchronous event-based data,”Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, Springer, 2020, pp. 136–152

2020
[10]

A Multi-Dimensional Covert Transaction Recognition Scheme for Blockchain,

Zhang, Z., Wang, S., Li, Z., Gao, F., and Wang, H., “A Multi-Dimensional Covert Transaction Recognition Scheme for Blockchain,”Mathematics, Vol. 11, No. 4, 2023, p. 1015

2023
[11]

Imagenet large scale visual recognition challenge,

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al., “Imagenet large scale visual recognition challenge,”International journal of computer vision, Vol. 115, 2015, pp. 211–252

2015
[12]

Evaluating collaborative filtering recommender systems,

Herlocker, J. L., Konstan, J. A., Terveen, L. G., and Riedl, J. T., “Evaluating collaborative filtering recommender systems,” ACM Transactions on Information Systems (TOIS), Vol. 22, No. 1, 2004, pp. 5–53

2004
[13]

The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets,

Saito, T., and Rehmsmeier, M., “The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets,”PloS one, Vol. 10, No. 3, 2015, p. e0118432

2015
[14]

The area under the precision-recall curve as a performance metric for rare binary events,

Sofaer, H. R., Hoeting, J. A., and Jarnevich, C. S., “The area under the precision-recall curve as a performance metric for rare binary events,”Methods in Ecology and Evolution, Vol. 10, No. 4, 2019, pp. 565–577

2019
[15]

Deep learning in news recommender systems: A comprehensive survey, challenges and future trends,

Talha, M. M., Khan, H. U., Iqbal, S., and Mohammed, “Deep learning in news recommender systems: A comprehensive survey, challenges and future trends,”Neurocomputing, Vol. 562, 2023, p. 126881. https://doi.org/https://doi.org/10.1016/j.neucom. 2023.126881, URL https://www.scien...

2023
[16]

Dropout: a simple way to prevent neural networks from overfitting,

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R., “Dropout: a simple way to prevent neural networks from overfitting,”The journal of machine learning research, Vol. 15, No. 1, 2014, pp. 1929–1958. 16

2014

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Retinomorphic event-based vision sensors: bioinspired cameras with spiking output,

Posch, C., Serrano-Gotarredona, T., Linares-Barranco, B., and Delbruck, T., “Retinomorphic event-based vision sensors: bioinspired cameras with spiking output,”Proceedings of the IEEE, Vol. 102, No. 10, 2014, pp. 1470–1484

work page 2014

[2] [2]

Collision detection for UAVs using event cameras,

Paulo, J. P. C., “Collision detection for UAVs using event cameras,” Ph.D. thesis, University of Lisbon, 2023

work page 2023

[3] [3]

Computer vision for autonomous vehicles: Problems, datasets and state of the art,

Janai, J., Güney, F., Behl, A., Geiger, A., et al., “Computer vision for autonomous vehicles: Problems, datasets and state of the art,”Foundations and Trends®in Computer Graphics and Vision, Vol. 12, No. 1–3, 2020, pp. 1–308

work page 2020

[4] [4]

Deep residual learning for image recognition,

He, K., Zhang, X., Ren, S., and Sun, J., “Deep residual learning for image recognition,”Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

work page 2016

[5] [5]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv preprint arXiv:2010.11929, 2020

arXiv 2010

[6] [6]

A large scale event-based detection dataset for automotive,

De Tournemire, P., Nitti, D., Perot, E., Migliore, D., and Sironi, A., “A large scale event-based detection dataset for automotive,” arXiv preprint arXiv:2001.08499, 2020

arXiv 2001

[7] [7]

Discussion on event-based cameras for dynamic obstacles recognition and detection for UAVs in outdoor environments,

Ben Miled, M., Zeng, Q., and Liu, Y., “Discussion on event-based cameras for dynamic obstacles recognition and detection for UAVs in outdoor environments,”UKRAS22 Conference" Robotics for Unconstrained Environments" Proceedings, EPSRC UK-RAS Network, 2022, pp. 34–35

work page 2022

[8] [8]

End-to-end learning of representations for asynchronous event-based data,

Gehrig, D., Loquercio, A., Derpanis, K. G., and Scaramuzza, D., “End-to-end learning of representations for asynchronous event-based data,”Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 5633–5643

work page 2019