acceptodds
Under review as a conference paper at ICLR 2027

QuEST: Quality-guided Evidence Selection and Temporal State Control for Event-based Object Detection

Abstract

Event activity reveals where changes occur, but its magnitude alone does not determine their relevance to object detection. Recurrent sparse detectors must therefore decide both which observations receive computation and how strongly they update temporal memory. We introduce QuEST (Quality-guided Evidence Selection and Temporal State Control), a lightweight controller that coordinates these decisions through a shared, detection-supervised quality field. Predicted from explicit event statistics, this field guides fixed-cardinality token selection, modulates selective state-space integration, and regulates cross-frame recurrent writing. Foreground-preserved consistency training regularizes predictions and control signals under changes in background event statistics, while retaining a single recurrent stream at inference. On the full Gen1 streaming test set, QuEST achieves 52.02 AP at 50% token retention, improving over an architecture-matched activity-only baseline by 1.30 AP and 1.88 AP₇₅ at identical per-stage scan cardinality, with 1.57% additional parameters and 6.02% additional backbone FLOPs. Without additional pretraining, QuEST also achieves 50.6 AP on 1Mpx and 33.2 AP on eTraM. Its AP advantage over the activity-only baseline persists across all five evaluated event-statistic shifts. These results support shared quality guidance as an effective criterion for coordinating sparse computation and temporal memory in event-based detection.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
event-based object detection, state space models, sparse computation, evidence selection, temporal modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_N2yEOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2(+1)
Soundness
1
Read review
Paper summary

This paper proposes QuEST, a quality-guided framework for event-based object detection that aims to improve the allocation of computational resources and temporal memory updates. Built upon SMamba, QuEST introduces a shared quality field estimated from event statistics to coordinate three components: Select, which guides sparse token selection; Scan, which modulates state-space integration; and Write, which regulates recurrent memory updates. Additionally, a foreground-preserved consistency training strategy is introduced to improve robustness against variations in background event statistics.

Critical strengths
  1. The proposed method integrates token selection, state-space integration, and recurrent memory updates through a shared quality field.

  2. The method is evaluated on three event-based object detection datasets, including Gen1, 1Mpx, and eTraM, and achieves consistent performance improvements

Critical weaknesses
  1. The research problem and motivation are not clearly articulated, and the analysis of existing challenges in the field lacks sufficient depth. As a result, it is difficult to identify the central focus and key contributions of the paper. The proposed method appears to be primarily an engineering-oriented extension of SMamba rather than a fundamentally novel methodological contribution. Moreover, the logical organization, coherence, and overall readability of the manuscript require substantial improvement

  2. The authors should provide comprehensive efficiency comparisons with existing methods, including computational complexity and inference latency.

  3. In the ablation studies, the performance improvements introduced by Scan and Write are relatively marginal. Are these gains consistent across different random seeds? The current experimental results are insufficient to convincingly establish the effectiveness of these components.

  4. Why is an additional Write gate introduced on top of the existing input and forget gates in ConvLSTM? Could similar effects be achieved by directly modulating the original ConvLSTM gates? Given the marginal performance improvement attributable to Write, stronger experimental evidence is needed to justify its necessity and effectiveness.

  5. The authors should compare the proposed quality-guided token selection strategy with existing approaches, such as SMamba and SAST, as well as more advanced visual token selection methods. Such comparisons are essential to demonstrate the advantages of the proposed selection mechanism over existing alternatives.

  6. All five evaluated perturbations are artificially constructed modifications of event statistics and closely resemble the perturbation types used during training. Even though foreground preservation is not applied during testing, the resulting distributions may still be similar to those covered by the training augmentations. More importantly, the baseline model is barely affected by these perturbations, suggesting that the current evaluation is insufficient to demonstrate the robustness and generalization capability of QuEST under realistic environmental shifts.

Questions for authors

See weaknesses

9 comments
AuthorsOpenReview ↗

Response to Reviewer N2yE (I/VI)

We sincerely thank the reviewer for taking the time to carefully evaluate our work and provide detailed feedback. The comments have encouraged us to further examine the motivation, methodological choices, and experimental evidence of QuEST.

(W1) The research problem and motivation are not clearly articulated, and the analysis of existing challenges in the field lacks sufficient depth. The proposed method appears to be primarily an engineering-oriented extension of SMamba rather than a fundamentally novel methodological contribution. Moreover, the logical organization, coherence, and overall readability of the manuscript require substantial improvement.

Reply 1: We thank the reviewer for pointing out the need to clarify the research problem and methodological contribution.

Our motivation starts from an important distinction in event-based detection: event activity indicates where changes occur, but does not necessarily indicate which observations are relevant to object detection. Existing sparse detectors, such as SAST and SMamba, have explored adaptive computation allocation through scene-adaptive sparsity and event-guided token selection. However, in recurrent detectors, spatial selection and temporal persistence are closely coupled. The tokens admitted for processing determine the evidence available to subsequent state updates, while recurrent writing determines how this evidence influences future intervals.

This motivates the central question of QuEST: under a fixed per-stage token budget, how should a detector coordinate which evidence is processed, how it is integrated, and how it is retained over time?

QuEST addresses this question through a shared, detection-supervised quality field derived from five explicit event statistics. It coordinates three operations:

  • Select: combines the activity prior with detection quality to refine Top-K token membership while maintaining fixed scan cardinality.
  • Scan: uses quality information to modulate state-space integration along the selected token sequence.
  • Write: regulates the incorporation of current evidence into recurrent hidden and cell states through quality-conditioned gating.

The contribution is therefore the coordination of spatial evidence selection, within-interval state integration, and cross-interval memory updates through a common detection-relevance signal, rather than simply adding independent modules to SMamba.

Our controlled ablations support this design. At 50% token retention, Select improves Gen1 AP from 50.72 to 51.14. Adding Scan increases AP to 51.43, and enabling Write further increases it to 51.86. Training regularization brings the final result to 52.02 AP.

We have also conducted additional experiments to examine these design choices. Independent quality heads achieve 52.0762 AP, compared with 52.0197 AP for the shared-field design, demonstrating comparable detection performance with a more unified controller. Density-only estimation achieves 51.0790 AP, 0.9407 points below full QuEST. Direct modulation of the original ConvLSTM gates achieves 51.7119 AP, compared with 51.86 AP for the proposed Write design. Further mechanism diagnostics show that the shared quality field independently influences Scan and Write, while temporal analyses reveal progressive recurrent-state and prediction recovery after temporary input perturbations are removed.

We also take the reviewer’s comment on presentation seriously. We will make the relationship between activity-based selection, detection relevance, and temporal memory more explicit in the Introduction, strengthen the comparison with SAST and SMamba in Related Work, and improve the transitions between quality estimation, Select, Scan, and Write in the Method section.

AuthorsOpenReview ↗

Response to Reviewer N2yE (II/VI)

(W2) The authors should provide comprehensive efficiency comparisons with existing methods, including computational complexity and inference latency.

Reply 2: We thank the reviewer for this suggestion. Our paper reports model parameters, backbone FLOPs, and inference runtime to evaluate both model complexity and practical inference cost.

We summarize the efficiency comparison with representative event-based detectors, including RVT, SAST, and SMamba:

MethodGen1 APParams (M)FLOPs (G)Runtime (ms)
RVT47.218.53.510.2
SAST48.218.92.422.7
SMamba50.416.12.424.8
QuEST (Ours)52.016.32.722.0

QuEST achieves the highest detection accuracy among the listed methods, with fewer parameters than RVT and SAST and a modest parameter increase over SMamba. Using QuEST’s AP of 52.02, the improvements over RVT and SAST are 4.82 and 3.82 AP points, respectively. Compared with SMamba, QuEST improves AP by 1.62 points, with a similar model size and a shorter measured runtime (22.05 ms vs. 24.84 ms).

We also provide a controlled comparison with our architecture-matched Activity-only baseline. Both models retain exactly 50% of tokens at every stage:

MetricActivity-onlyQuEST
AP50.7252.02
AP7552.7654.64
Parameters (M)16.07016.322
Backbone FLOPs (G)2.5762.731
Runtime (ms)20.5122.05
Token retention50%50%

The 1.30-point AP improvement comes with 1.57% additional parameters and 6.02% additional backbone FLOPs, while runtime increases by 1.54 ms. The number of tokens processed by selective scanning remains unchanged; the additional computation comes from the quality-guided control components.

Together, these comparisons demonstrate improved detection performance at identical per-stage scan cardinality, with modest increases in model size and inference cost relative to the architecture-matched Activity-only baseline.

AuthorsOpenReview ↗

Response to Reviewer N2yE (III/VI)

(W3) In the ablation studies, the performance improvements introduced by Scan and Write are relatively marginal. Are these gains consistent across different random seeds? The current experimental results are insufficient to convincingly establish the effectiveness of these components.

Reply 3: We appreciate the reviewer’s concern regarding training variability. Our controlled ablations show positive incremental gains from both components. With the backbone, neck, head, quality supervision, and per-stage token retention held fixed, adding Scan improves Gen1 AP from 51.14 to 51.43, and subsequently enabling Write increases AP to 51.86. Together, Scan and Write provide a 0.72-point AP improvement over quality-guided selection alone.

Our implementation follows the recurrent streaming training and evaluation framework established by RVT, which is also adopted by subsequent event-based detectors such as SAST and SMamba. In this setting, training combines streaming sequence sampling and randomly sampled sequences, with recurrent states and data ordering handled across successive event intervals. Reproducibility through conventional global seed control is not straightforward. In particular, the RVT implementation explicitly disables PyTorch Lightning’s global seed setting because of unresolved issues with shuffling during streaming-dataset training, as reflected in the training log in RVT Issue #39.

The published RVT, SAST, and SMamba results also report individual benchmark scores rather than systematic mean ± standard deviation results over multiple independent training seeds. Following this established training and evaluation protocol, we report our controlled ablations under the same setting. These results quantify component gains under the adopted protocol; they do not quantify variability across independent training runs.

To complement the detection results, we further examined how Scan and Write use the quality signal to influence state processing. With token selection unchanged, independently replacing the quality inputs to Scan and Write with constant or spatially shuffled maps produces measurable changes in downstream features and recurrent states. These interventions show that quality affects state processing beyond its role in token selection.

For Scan, inspection of the actual selective-scan tensors confirms that quality modulates the effective integration steps and state-transition behavior. For Write, the full-test-set spatial analysis shows higher average write rates in foreground regions, indicating a stronger contribution from candidate states in object-related regions. Temporal recovery analysis further shows that hidden-state and cell-state deviations progressively decrease after temporary perturbations are removed, accompanied by improved prediction consistency.

Together, the controlled AP comparisons and mechanism-level analyses support the complementary contributions of Scan and Write within the evaluated framework: Scan regulates integration along the selected token sequence, while Write controls the incorporation of candidate states into recurrent memory.

Reviewer_eDpcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes QuEST (Quality-guided Evidence Selection and Temporal State Control), a framework for event-based object detection that uses a shared, detection-supervised quality field to jointly guide sparse token selection and temporal state updates. The authors also introduce foreground-preserved consistency training to improve robustness against background event variations.

Experiments on Gen1, 1Mpx, and eTraM demonstrate consistent improvements over SMamba. On Gen1, QuEST achieves 52.02 AP with 50% token retention, outperforming an architecture-matched activity-only baseline by 1.30 AP with modest computational overhead.

Critical strengths
  1. Clear Motivation and Well-Defined Problem

The paper addresses an important limitation of existing event-based object detection methods: high event activity does not necessarily indicate high detection relevance. The motivation is clear, and the proposed approach directly addresses this issue.

  1. Unified Sparse Selection and Temporal State Control

The paper introduces a unified framework that uses a shared quality field to guide both sparse token selection and temporal state updates. This design effectively connects spatial computation with temporal memory management.

  1. Fair Ablation Studies and Good Performance-Efficiency Trade-off

The ablation studies maintain the same backbone architecture and token budget, enabling fair comparisons between different components. The proposed method achieves consistent accuracy improvements with relatively small increases in parameters and computational cost.

Critical weaknesses
  1. Limited Novelty in Individual Components

The technical novelty mainly lies in integrating existing techniques rather than introducing fundamentally new mechanisms. The individual components, such as quality-guided token selection and state gating, are relatively straightforward.

  1. Insufficient Justification for the Shared Quality Field

The paper does not sufficiently explain why token selection, state-space integration, and memory writing should share the same quality field. Since these operations have different objectives, separate quality signals might be more effective.

  1. Potential Memory Inertia Introduced by the Write Module

The additional Write gate may introduce redundant control, as ConvLSTM already contains mechanisms for updating and retaining memory. This could make the model overly conservative when responding to sudden scene changes or newly appearing objects.

  1. Limited Control over SSM Dynamics in the Scan Module

The Scan module only adjusts the SSM integration step size based on quality scores, which provides limited flexibility in controlling state dynamics. The paper also lacks sufficient analysis of how this adjustment affects temporal information modeling.

Questions for authors
  1. Could the authors provide a more detailed ablation study of the prediction, feature, quality, gate consistency, and environment alignment losses? It would be helpful to understand which component contributes most to the substantial improvement under nonnegative noise perturbations.

  2. Since ConvLSTM already has input and forget gates, could the authors further clarify the benefits of introducing an additional quality-conditioned write gate? I also wonder whether this extra mechanism might lead to slower memory updates, especially during sudden scene changes or when new objects appear.

  3. Could the authors provide more insights into the contributions of the five event-statistic channels to quality estimation? It would also be interesting to see whether a simpler density-based estimator with the same supervision could achieve similar performance.

  4. Could the authors further discuss how the quality-guided recurrent writing mechanism handles challenging temporal scenarios, such as sudden motion changes, objects entering or leaving the scene, and occlusions?

6 comments
AuthorsOpenReview ↗

Response to Reviewer eDpc (I/IV)

We sincerely thank the reviewer for the thoughtful and constructive feedback. The reviewer’s comments have helped us further examine the design choices of QuEST and improve the clarity and completeness of our work. In response, we have conducted additional experiments and mechanism-level analyses, including comparisons with independent quality heads, alternative ConvLSTM gate modulation, and density-only quality estimation, together with further investigations into temporal state recovery, SSM dynamics, and the contributions of individual event statistics. Given the limited time and GPU resources during the rebuttal period, fully training and evaluating additional model variants requires considerable computational time. Nevertheless, we have made every effort to provide further experimental evidence using both newly trained models and existing checkpoints. Our point-by-point responses are presented below.

(W)1. Limited Novelty in Individual Components.

(W)2. Insufficient Justification for the Shared Quality Field.

Reply 1: We thank the reviewer for raising these important questions.

Although Select, Scan, and Write have different operational objectives, they all depend on a common underlying criterion: the detection relevance of the available evidence. Our motivation is to use a shared, detection-supervised quality field to provide this common signal, while allowing each operation to translate it into its own control behavior. Select uses quality to refine token membership, Scan uses it to modulate state-space integration, and Write uses it to regulate recurrent updates.

To examine whether separate quality predictors could be more effective, we trained an Independent Heads variant. The complete Gen1 results are:

VariantAPAP50AP75AP_SAP_MAP_L
QuEST (shared)52.019780.930854.636844.062058.464747.9880
Independent Heads52.076280.240155.795046.660656.052160.6214

Independent Heads achieves a slightly higher overall AP, with a difference of only 0.0565 points. The two designs exhibit comparable overall detection performance, with different results across object-size categories. This comparison shows that a single quality field can coordinate the three operations without requiring separate quality prediction heads, providing a more unified and compact control design while retaining comparable detection performance.

We further examined whether the shared quality field actually influences Scan and Write. By independently replacing the quality input to each module with spatially constant or shuffled values while keeping token selection unchanged, we obtained the following results:

InterventionFeature rel. L1Hidden rel. L1Cell rel. L1Gate-map MAE
Constant Scan0.01680.01680.01560.0024
Shuffled Scan0.02240.02240.02010.0030
Constant Write0.10200.10200.13620.0865
Shuffled Write0.12370.12370.16470.1062

Spatially shuffling quality increases downstream feature deviations in both modules, by 33.0% for Scan and 21.2% for Write, compared with their respective constant-quality controls. Since token selection remains unchanged, these differences directly reflect the influence of quality on subsequent state processing and recurrent writing.

We believe the contribution of QuEST lies in coordinating evidence selection, within-interval state integration, and cross-interval memory updates through a shared detection-relevance criterion, rather than simply combining three independent modules.

AuthorsOpenReview ↗

Response to Reviewer eDpc (II/IV)

3. Potential Memory Inertia Introduced by the Write Module.

Q2. Since ConvLSTM already has input and forget gates, could the authors further clarify the benefits of introducing an additional quality-conditioned write gate? I also wonder whether this extra mechanism might lead to slower memory updates, especially during sudden scene changes or when new objects appear.

Q4. Could the authors further discuss how the quality-guided recurrent writing mechanism handles challenging temporal scenarios, such as sudden motion changes, objects entering or leaving the scene, and occlusions?

Reply 2: We thank the reviewer for these questions. The purpose of Write is to incorporate detection-relevance information into recurrent state updates, regulating how much of the current candidate state is integrated with the previously accumulated memory.

Comparison with direct ConvLSTM gate modulation.

To examine whether directly modulating the original ConvLSTM gates could achieve similar results, we implemented a Direct Gate alternative. The complete Gen1 results are:

VariantAPAP50AP75AP_SAP_MAP_L
Select + Scan51.430080.650054.300043.700057.940047.1400
Direct Gate51.711980.070855.443045.864756.386460.9912
QuEST w/o Reg.51.860081.040054.140044.480058.060052.2500
QuEST52.019780.930854.636844.062058.464747.9880

The proposed Write design achieves 51.86 AP, compared with 51.71 AP for direct gate modulation. This result supports the effectiveness of explicitly regulating candidate-state incorporation through a quality-conditioned write rate.

Spatially selective recurrent writing.

We further analyzed the spatial distribution of Write rates over the full Gen1 test set, covering all 470 test sequences and 30,605 labeled frames. Foreground and background write rates are calculated as conditional regional means, with each region normalized by its own number of spatial locations.

StageFG UpdateBG UpdateFG-BG Difference
10.59840.4770+0.1215
20.52950.4244+0.1051
30.45140.4209+0.0305
40.44390.4327+0.0112

Foreground regions consistently receive higher write rates than background regions across all four stages, with stronger differences in the earlier stages. Since the write gate controls the mixing of previous and candidate recurrent states, higher foreground write rates indicate a stronger contribution from current candidate evidence in object-related regions. This demonstrates that Write performs spatially adaptive memory updates rather than applying a uniform recurrent update.

Temporal response and state recovery.

To examine potential memory inertia, we injected noise during frames 4–11 and restored clean inputs from frame 12 onward. We tracked the hidden states, cell states, and Write gates along both recurrent streams:

FrameHidden rel. L1Cell rel. L1Gate-map MAE
11 (last perturbed)0.16000.16320.0707
120.12020.14080.0093
140.08940.11720.0081
160.07410.10280.0074
180.06440.09240.0067
200.05730.08430.0063

After the perturbation is removed, hidden-state and cell-state deviations decrease by 64.2% and 48.4%, respectively, while the Write-gate deviation decreases by 91.1%.

The corresponding detection outputs also recover. At frame 20, the raw prediction relative L1 decreases to 0.0058, confidence-list MAE to 0.0061, and clean-to-perturbed box IoU reaches 0.9761.

Target entry and exit analysis.

We additionally examined selected Gen1 streaming clips around target-presence transitions:

Event typeClipsObserved result
Target entry10First matched detection latency: 0 frames
Target exit10Mean residual prediction duration: 2.3 frames
Motion-change heuristic10Mean quality change: +0.0009
Motion-change heuristic10Mean gate change: -0.0127
Occlusion gap0No qualifying clips identified

The target-entry and target-exit clips provide additional observations of recurrent behavior around changes in target presence. For occlusion, the label-presence scan did not identify qualifying gaps followed by reappearance, so no quantitative occlusion result is reported.

Together, these experiments show that the proposed Write mechanism provides spatially adaptive memory updates and exhibits progressive state and prediction recovery after temporary input perturbations.

AuthorsOpenReview ↗

Response to Reviewer eDpc (III/IV)

(W)4. Limited Control over SSM Dynamics in the Scan Module.

Q3. Could the authors provide more insights into the contributions of the five event-statistic channels to quality estimation? It would also be interesting to see whether a simpler density-based estimator with the same supervision could achieve similar performance.

Reply 3: We thank the reviewer for this suggestion. Our quality estimator uses five event statistics: density (D), temporal entropy (E), polarity imbalance (P), temporal center (C), and occupancy (O). These statistics describe complementary aspects of event distributions beyond their overall magnitude.

Density-only comparison.

To examine whether density alone is sufficient, we trained a Density-only variant using the same quality supervision and control framework, but with only the density channel as input. The complete Gen1 results are:

VariantAPAP50AP75AP_SAP_MAP_L
Activity-only50.720080.630452.758043.019357.199950.1400
Density-only51.079079.202154.466044.627556.308153.6744
QuEST52.019780.930854.636844.062058.464747.9880

Density-only improves over Activity-only by 0.3590 AP, showing that supervised density information provides useful guidance. However, it remains 0.9407 AP below the full five-statistic estimator. The full estimator also achieves higher AP50 and AP_M, while Density-only obtains higher AP_S and AP_L.

Five-channel intervention analysis.

We further performed inference-time interventions on the five channels of the trained full model. Each intervention changes only the quality encoder input while keeping the checkpoint and subsequent control framework unchanged.

InterventionQuality MAEQuality Corr.Selection JaccardFeature rel. L1
Zero-D0.03970.99360.98730.0538
Zero-E0.03490.99240.98760.0598
Zero-P0.04930.99000.98820.0895
Zero-C0.03140.99690.99140.0400
Zero-O0.21100.15840.89820.2177
Density-only input0.07990.89570.95860.1517

Among the five channels, occupancy has the strongest influence on the learned quality field. Removing it causes the largest changes in quality prediction, token selection, and downstream features.

Spatial shuffling provides complementary evidence. In particular, shuffling the occupancy channel reduces quality correlation to 0.8294 and selection Jaccard to 0.9443, while increasing feature deviation to 0.1266. These interventions show that both the values and spatial correspondence of event statistics contribute to the learned quality-control behavior.

Quality-conditioned SSM dynamics.

We also inspected the actual selective-scan tensors to understand how quality affects state-space integration. The effective integration step is:

Δeff=Δbase(0.5+0.5q).\Delta_{\mathrm{eff}} = \Delta_{\mathrm{base}}(0.5+0.5q).

For a negative diagonal transition matrix AA, the discrete transition is:

A‾=exp⁡(AΔeff).\overline{A} = \exp(A\Delta_{\mathrm{eff}}).

In Stage 1 under clean inputs, increasing quality from approximately 0.375 to 0.822 increases the effective integration step from 0.0178 to 0.0247, while the transition factor decreases from 0.9836 to 0.9780. A similar trend is observed in Stage 3, where the effective step increases from 0.0457 to 0.0730 and the transition factor decreases from 0.9642 to 0.9484.

Thus, higher-quality evidence receives a larger integration step, allowing the current input to exert a stronger influence on the SSM state. Lower-quality evidence uses a smaller step, retaining more of the preceding scan-state contribution. This analysis directly illustrates how the shared quality field modulates the state-space dynamics.

Reviewer_x8NQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

QuEST proposes an SMamba-based event-based object detection method. It introduces a quality field to control fixed-cardinality token selection, state-space scanning fusion, and other operations. The method is validated on the Gen1, 1Mpx and eTraM datasets and achieves promising experimental results.

Critical strengths
  1. The proposed method is evaluated on multiple datasets and outperforms the baseline methods. Comprehensive performance results are also reported, including the number of parameters and specific detection metrics.
  2. The ablation experiments in this paper are comprehensive, illustrating the performance gains brought by each component and verifying the benefits of feature selection, scan control, and other modules.
Critical weaknesses
  1. Although the proposed method achieves certain performance gains over the baseline, such gains are not significant and introduce increased computational complexity to some extent, making the improvements rather limited.

  2. This work is implemented based on state-space methods for event-based object detection. This is a relatively well-studied problem in the event vision field, and the adopted architecture is also classic. I would prefer to see classic approaches applied to new problems or novel methods developed to tackle established problems.

  3. All comparative experiments adopt a single SMamba/ConvLSTM/YOLOX architecture. It remains unclear whether this shared quality controller can generalize to other sparse backbone networks, and it is also undetermined whether its performance benefits rely on this specific implementation.

Questions for authors

I hope the authors can clearly elaborate on the second issue I listed under Critical Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer x8NQ (I/III)

We sincerely thank the reviewer for the thoughtful and constructive feedback, as well as for recognizing the multi-dataset evaluation and ablation studies. The comments have encouraged us to further clarify the motivation and contribution of QuEST. Our point-by-point responses are presented below.

(W)1. Although the proposed method achieves certain performance gains over the baseline, such gains are not significant and introduce increased computational complexity to some extent, making the improvements rather limited.

Reply 1: We agree that detection accuracy should be considered together with computational cost. Our objective is not to improve accuracy simply by increasing computation, but to make better use of a fixed sparse scanning budget.

On Gen1, QuEST achieves 52.02 AP, improving over the architecture-matched Activity-only baseline by 1.30 AP and 1.88 AP75, while maintaining the same 50% token retention at every stage. This improvement requires only 1.57% more parameters and 6.02% additional backbone FLOPs. The measured inference time increases from 20.51 ms to 22.05 ms. Thus, the accuracy improvement is achieved without increasing the number of scanned tokens.

We also compare QuEST with methods employing additional training or adaptation. As reported in Table 1, EvRT-DETR-B and TESPEC achieve 52.7 and 51.6 AP on Gen1, respectively, but use 57.1M and 46.7M parameters, compared with only 16.3M for QuEST. EvRT-DETR-B employs staged adaptation, while TESPEC uses self-supervised pretraining. QuEST achieves competitive detection performance without additional pretraining. On 1Mpx, QuEST reaches 50.6 AP, matching TESPEC and surpassing EvRT-DETR-B (50.1 AP), while achieving 33.2 AP on eTraM.

Furthermore, our controlled ablations demonstrate complementary contributions from Select, Scan, and Write. We also evaluated alternative designs, including independent quality heads, density-only quality estimation, and direct ConvLSTM gate modulation. Together with the additional mechanism diagnostics, these experiments provide further evidence for the design choices of QuEST.

Overall, we believe the results demonstrate a favorable balance between detection accuracy, model size, and computational cost.

AuthorsOpenReview ↗

Response to Reviewer x8NQ (II/III)

(W)2. This work is implemented based on state-space methods for event-based object detection. This is a relatively well-studied problem in the event vision field, and the adopted architecture is also classic. I would prefer to see classic approaches applied to new problems or novel methods developed to tackle established problems.

Reply 2: We appreciate the reviewer's perspective that applying established methods to new problems is a valuable research direction. In this work, we take a complementary direction by developing a new approach to an existing but less explored problem in event-based object detection.

Our motivation comes from the observation that event activity indicates where changes occur, but does not necessarily reflect their relevance to object detection. Existing sparse detectors have explored adaptive computation from different perspectives. For example, SAST adapts window and token sparsity to scene activity, while SMamba combines event-guided sparsification with state-space scanning. QuEST addresses a complementary question: under a fixed scan budget, how can a detector coordinate which evidence is processed with how that evidence is integrated and retained over time?

This question arises from the close relationship between spatial selection and temporal persistence in recurrent detectors. The selected tokens determine what evidence is available for subsequent state updates, while recurrent writing determines how this evidence influences future intervals.

To address this problem, QuEST introduces a shared, detection-supervised quality field derived from five event statistics: density, temporal entropy, polarity imbalance, temporal center, and occupancy. This field coordinates three operations under a fixed per-stage token budget:

  • Select combines the activity prior with detection quality to refine Top-K token membership without increasing the number of selected tokens.
  • Scan uses quality information to modulate state-space integration along the selected token sequence.
  • Write uses quality-conditioned gating to regulate how current evidence is incorporated into recurrent hidden and cell states.

The central idea is to use a common detection-relevance signal to coordinate evidence selection, within-interval state integration, and cross-interval memory updates. In this way, QuEST connects spatial computation allocation with temporal memory management rather than treating them as independent decisions.

Our controlled ablations support this design. At the same 50% token retention, Select improves Gen1 AP from 50.72 to 51.14. Adding Scan further increases AP to 51.43, and enabling Write brings it to 51.86. Training regularization further improves the final result to 52.02 AP.

We further conducted several experiments to examine these design choices, including comparisons with independent quality heads, density-only estimation, and direct ConvLSTM gate modulation. These experiments provide additional evidence for the shared quality field, the use of multiple event statistics, and the proposed Write mechanism. The corresponding quantitative comparisons and detailed mechanism analyses are provided in our responses to Reviewer eDpc. In particular, the Scan/Write interventions and temporal diagnostics further demonstrate how the shared quality field influences state-space integration and recurrent memory updates.

QuEST builds upon established components, including SMamba, ConvLSTM, and YOLOX. Its contribution lies in using detection-supervised quality to coordinate these components under a fixed per-stage token budget. We believe that addressing the relationship between sparse evidence selection and temporal memory provides a meaningful methodological contribution to event-based object detection.

AuthorsOpenReview ↗

Response to Reviewer x8NQ (III/III)

(W)3. All comparative experiments adopt a single SMamba/ConvLSTM/YOLOX architecture. It remains unclear whether this shared quality controller can generalize to other sparse backbone networks, and it is also undetermined whether its performance benefits rely on this specific implementation.

Reply 3: We thank the reviewer for highlighting the potential of extending this approach to other sparse architectures. We agree that the current evaluation is based on the SMamba detection framework and does not yet establish generalization across different backbone families.

To describe the architecture, QuEST retains SMamba's four-stage sparse backbone: Stages 1–2 use Sparse Spatial Mamba blocks, while Stages 3–4 use Spatial-Channel Mixing Mamba (SCMM) blocks. ConvLSTM maintains recurrent states at each stage, and the YOLOX neck and head perform multi-scale feature aggregation and object prediction. We keep these components unchanged throughout our experiments to isolate the effects of the proposed quality-guided controls.

The proposed quality-guided evidence selection principle is not conceptually restricted to SSM backbones. For Transformer-based sparse event detectors such as SAST, the quality field could guide token admission under a fixed per-stage token budget. The Scan mechanism, which currently modulates SSM integration, would need to be adapted to Transformer-based feature processing. Similarly, Write could be adapted to the corresponding temporal memory structure.

We are particularly interested in extending quality-guided selection to Transformer-based sparse backbones. Such cross-architecture evaluation would require adapting the control mechanisms and conducting matched training and evaluation, providing a natural direction for further development of QuEST.

In our current experiments, retaining the same backbone, neck, and detection head allows us to make controlled comparisons and attribute the observed improvements to quality-guided evidence selection and temporal state control. The consistent results on Gen1, 1Mpx, and eTraM, together with the component ablations and additional mechanism analyses, demonstrate the effectiveness of this design within the evaluated recurrent sparse SSM detector.

We sincerely thank the reviewer for taking the time to carefully evaluate our work and provide thoughtful and constructive suggestions. The feedback has been valuable in helping us clarify the motivation, strengthen the experimental evidence, and improve the presentation of QuEST. We greatly appreciate the reviewer’s time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.