Talk2Event: Grounded Understanding of Dynamic Scenes from Event Cameras

We sincerely thank Reviewer botn for the thoughtful evaluation, positive feedback on our contributions and writing clarity, and for highlighting both the novelty of Talk2Event and the aesthetic quality of our visualizations. Below, we address the weaknesses you raised in detail.

---

Q1: "The proposed dataset is based on DSEC, which is not very rich in scene diversity. The train set and test set are quite similar, which makes the in-the-wild generalization ability of the proposed method a concern."

A: Thank you for pointing this out. We acknowledge that the DSEC dataset, while high-quality and well-suited for event camera research, might have limited diversity in geographic locations and traffic conditions.

To mitigate this:

• We curated Talk2Event with a strict split-by-sequence protocol, ensuring zero overlap in trajectories between the training and test sets.
• We emphasized intra-sequence diversity by including scenes under varying lighting, weather, and motion conditions (e.g., fast-moving vehicles, crosswalks, occlusions).
• Our model is tested across event-only, frame-only, and fusion modalities, further challenging generalization under partial observation settings.

We agree that broader generalization remains an important direction. In future iterations, we plan to extend Talk2Event to more diverse sources (e.g., M3ED) and to cross-domain evaluation settings. This discussion has been added in Section 6 in the manuscript.

---

Q2: "Annotation and evaluation are conducted over only 7 object classes. More classes would lead to a further step to open-vocabulary grounding."

A: We fully agree. Our current focus on 7 foreground classes (car, truck, bus, pedestrian, bicycle, motorcycle, and rider) reflects the annotation scope available in DSEC. We chose these classes to ensure high-quality and consistent box annotations from the underlying dataset.

Specifically:

• The referring expressions themselves are open-ended and include diverse natural language beyond the object category (e.g., behavior, position, and visual appearance).
• Our method is compatible with open-vocabulary grounding: the text encoder is pretrained with large-scale language supervision, and grounding is guided by the semantics of expressions rather than fixed class labels.

As suggested, we are exploring extensions toward zero-shot and long-tail categories using models pretrained on broader datasets (e.g., SAM or GroundingDINO-style detectors). We have added this plan to our future work section.

---

Q3: "The analysis of the 'Effects of Different Attributes' ablation study is confusing. The setting [(a), (s), (v), (o)] may not align with the conclusion that, for example, 'status' improves over 'appearance'."

A: Thank you for this important observation. We agree that the phrasing in our ablation analysis might have introduced confusion. Our current experiment isolates each attribute individually, i.e., comparing models that use only one of the four attribute tokens at a time.

We agree that a more insightful analysis would involve drop-one or leave-one-out combinations, such as $(a + s + v)$, $(a + s + o)$, etc. We have now:

• Clarified the intention of the original ablation (isolated expert performance).
• Revised Figure 5 and its caption to improve interpretability, aligning with your suggestion.

Based on your suggestions, we will add new results based on combinatorial attribute configurations, which better support conclusions on relative importance and complementarity. Due to time constraint, these new results will be included in Section C.1 of the revised supplementary material. Thank you again for helping improve the rigor of our analysis.

---

Q4: "How does the mixture-of-experts structure benefit the model performance? I see that explainability is improved by manually designing four experts and assigning to them different attributes to focus on, but could end-to-end training on all tokens bring better results?"

A: Thank you for the thoughtful question.

To clarify: our MoEE module is trained in a fully end-to-end manner. While each expert is designed to specialize in a grounding attribute (Appearance, Status, Relation-to-Viewer, Relation-to-Others), the weighting across experts is dynamically learned through a soft attention mechanism based on scene and expression context.

This design allows the model to:

• Adaptively attend to the most relevant attributes per sample, without enforcing hard separation or prior weighting.
• Retain compositional interpretability, as expert activations can be visualized and attributed to specific reasoning paths (e.g., Fig. 5 and Fig. 6).
• Achieve strong quantitative performance under diverse modality settings, as shown in the ablation results (Table 4).

In contrast to a monolithic design trained over all tokens, our framework strikes a balance between structured inductive bias and flexible, learnable fusion, enabling better grounding in dynamic and attribute-rich environments.

As suggested, we have further revised Section 4.2 in the manuscript to make the end-to-end nature of MoEE learning more explicit.

---

Q5: "State-of-the-art multimodal large language models (such as ChatGPT) also have the ability of visual grounding. How is their performance on this dataset?"

A: Thanks for your question. Multimodal LLMs like ChatGPT (via GPT-4V) or Gemini-1.5 have demonstrated impressive image-text reasoning capabilities. However, applying them to event-based visual grounding presents several limitations:

• Event data requires specialized encoding (e.g., voxelization, polarity channels), which is currently unsupported by existing VLM APIs.
• These models are not designed to predict bounding boxes; their outputs are primarily free-form text or coarse heatmaps at best.

We believe these models can be evaluated on high-level QA tasks (e.g., describing scenes or answering referring questions). We are currently conducting such experiments and plan to release a language-based diagnostic split of Talk2Event for this purpose.

We agree with you that comparing against foundation models is meaningful, not necessarily as direct competitors, but as complementary baselines. A discussion of this potential is included in our revised Section 6.

---

Q6: "The small visual icons in Table 1 are cute, and the icons for Frame, RGB-D, and LiDAR are self-explanatory. However, I cannot see the connection between the Event icon and its features. It would be nice if this could be pointed out."

A: Thank you for the kind comment and helpful suggestion. The “Sensory Data” column in Table 1 is designed to highlight the diversity of sensory modalities supported by each grounding benchmark, as mentioned in the table caption.

While prior datasets mostly focus on traditional RGB images or 3D point clouds (e.g., RGB-D, LiDAR), Talk2Event is the first to include event-based perception for visual grounding, making it a novel addition in this dimension.

For better clarity, we have further refined the caption based on your suggestion.

---

Reference:

• [M3ED] Chaney, K., Cladera, F., Wang, Z., Bisulco, A., Hsieh, M. A., Korpela, C., ... & Daniilidis, K. (2023). "M3ED: Multi-Robot, Multi-Sensor, Multi-Environment Event Dataset." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (pp. 4016-4023).

---

Last but not least, we would like to sincerely thank Reviewer botn again for the valuable time and constructive feedback provided during this review.

We sincerely thank Reviewer TxK4 for the thoughtful and encouraging feedback. We are glad you found the proposed benchmark and framework valuable and impactful. Below, we address your specific concerns in detail.

---

Q1: "The description of MoEE is too brief in the main text, making it difficult to fully understand without referring to the appendix."

A: Thank you for pointing this out. Due to page limits, we did not include too many details on MoEE in the main text. We agree that the MoEE (Mixture of Event-Attribute Experts) plays a central role in our model, and its workings deserve clearer exposition in the main text. To address this:

• We have revised Section 4.2 to include explicit formulae and an expanded explanation of the gating mechanism.
• Fig. 3 has been enhanced with better module labeling and input/output dimensions to make the information flow more intuitive.
• For completeness, we retain the step-by-step breakdown and intermediate shape definitions in the supplementary material (Section B.2).

These revisions aim to help readers grasp the operation and motivation of MoEE without needing to dive into supplementary details.

---

Q2: "Dataset generation depends heavily on a single vision-language model (Qwen2-VL), raising concerns about linguistic bias."

A: This is a valid and important concern. To reduce potential linguistic bias from over-reliance on Qwen2-VL, we took several key steps:

• We implemented attribute-conditioned prompting to diversify the structure and semantics of the generated expressions (more details in Section 3.2).
• The generated expressions were passed through multiple post-processing stages, including:
  • Redundancy filtering via lexical similarity thresholds;
  • Visibility checks with object masks and occlusion priors;
  • Manual spot-checking for semantic fluency and grounding validity.

In addition, the final dataset includes a broad variety of phrasings (e.g., varying determiners, active vs. passive voice, and presence or absence of attributes). As shown in our word clouds and examples (Fig. 2 in the manuscript, and Section A.4 in the supplementary material), the expressions demonstrate natural variation.

We acknowledge that using additional generation models (e.g., Gemini, GPT-4V) could further enhance stylistic diversity. We plan to incorporate this in a future version of Talk2Event and have discussed this point in Section 2 of the revised manuscript.

---

Q3: "Evaluation lacks metrics that specifically assess attribute-aware and relational reasoning."

A: Thank you for highlighting this important direction. Our current metrics (Top-1 Accuracy, mIoU) are standard in visual grounding and provide spatial accuracy assessment. However, we agree that they do not fully capture attribute-level or relational understanding.

To partially address this, we include:

• Expert activation heatmaps (see Fig. 5 and Fig. 6 in the manuscript) that visualize how the model attends to different attribute types across diverse scenes;
• Attribute-level ablation study (see Table 4 in the manuscript) that quantifies the contribution of each attribute module to performance.

We agree that introducing attribute-specific evaluation metrics would provide more direct insight into compositional reasoning. Inspired by semantic segmentation evaluation, we are currently exploring per-attribute correctness tracking (e.g., AP or IoU conditioned on attribute presence). We have added this discussion as a future benchmark direction in the revised Section 5.

---

Q4: "Lack of discussion on model efficiency such as latency, parameter count, or suitability for real-time applications."

A: We appreciate this feedback and agree that understanding computational efficiency is essential, especially for deployment in latency-critical systems like robotics and autonomous driving.

In response, we have included the following details in Section 5 of the revised manuscript and Section B.2 in the supplementary material:

• Parameter count: EventRefer has approximately 60M parameters, comparable to lightweight DETR-style models.
• Inference speed: On an NVIDIA RTX 3090 GPU, EventRefer runs at ~38 ms per frame (i.e., ~26 FPS), including event voxelization, feature encoding, and grounding.
• Memory usage: Peak GPU memory during inference is <5.2 GB, making it suitable for mid-range hardware.

These measurements demonstrate that the model is feasible for real-time use cases and supports further optimization (e.g., through model distillation or token pruning). We thank the reviewer for encouraging us to clarify this aspect.

---

Q5: "How were the four attributes (Appearance, Status, Relation-to-Viewer, Relation-to-Others) selected as core categories for annotation? Were they supported by data analysis or a preliminary study?"

A: Thank you for this thoughtful question. The four attributes were selected through an iterative design process combining linguistic analysis, pilot user study, and task requirements of real-world grounding in driving scenarios.

Specifically:

• We first reviewed prior visual grounding datasets (e.g., RefCOCO, ScanRefer, M3DRefer) and found that a large proportion of referring expressions involved descriptions of appearance and status, as well as spatial/relational cues.
• We then conducted a manual annotation pass over ~500 scene-expression pairs from a preliminary event dataset. We observed that over 94% of expressions could be semantically decomposed into some combination of the four proposed dimensions.
• We validated this through an internal study, where annotators independently labeled attributes per expression with high inter-rater agreement.

These four attributes thus emerged as both frequent and discriminative categories that align with how humans refer to objects in dynamic environments. For better clarity, we included further statistics and examples in Section A.2 and Section A.3 in the supplementary material.

---

Q6: "Can EventRefer detect invalid or ambiguous queries (e.g., 'this car' when multiple cars exist)? Was handling this considered in the design?"

A: This is a highly relevant question for practical deployment. While our current version of EventRefer does not include an explicit invalid-query rejection module, we designed the model to be sensitive to ambiguity through several mechanisms:

• The fuzzy matching step produces soft attention maps over ambiguous tokens, which often result in low-activation expert weights if the attributes are vague or under-specified.
• The scoring head outputs a confidence score, which can serve as a proxy for detection certainty. Ambiguous expressions typically yield lower confidence and higher entropy in attention distributions.

Although not the primary focus of this paper, we fully agree with you that incorporating ambiguity detection (e.g., via uncertainty estimation, query entropy, or calibration modules) would improve robustness. We are exploring this in follow-up work and have added this discussion in the revised future directions.

---

Q7: "How does the model behave under ambiguous or invalid queries? Can it abstain or indicate low confidence?"

A: Thank you for the follow-up. In its current form, EventRefer always outputs a box prediction, as is standard in most grounding benchmarks. However, we have observed in our internal experiments that:

• For ambiguous queries (e.g., “the car”), the model often exhibits low activation in all expert branches and generates low-confidence box scores.
• These scores could be used to threshold predictions or trigger fallback strategies (e.g., asking for clarification in interactive settings).

While abstention is not natively supported, we believe that combining the output confidence with a learned rejection head or training on ambiguous/negative samples (e.g., using an auxiliary “no-object” token) would enable the model to indicate “cannot locate” predictions.

We have now discussed this important capability as a design extension in the revised Section 6. Thanks again for your valuable suggestion.

---

Q8: "Would the authors consider supplementing with an efficiency analysis to better assess practical feasibility?"

A: Absolutely. As also mentioned in our response to Q4, we now include a full efficiency profile in the revised manuscript (Section 5) and Appendix Section B.2.

Key highlights:

• Parameter count: EventRefer has approximately 60M parameters, comparable to lightweight DETR-style models.
• Inference speed: On an NVIDIA RTX 3090 GPU, EventRefer runs at ~38 ms per frame (i.e., ~26 FPS), including event voxelization, feature encoding, and grounding.
• Memory usage: Peak GPU memory during inference is <5.2 GB, making it suitable for mid-range hardware.

These numbers confirm that EventRefer is lightweight and deployable in real-time settings, which we believe strengthens the practical value of the proposed method.

---

Last but not least, we would like to sincerely thank Reviewer TxK4 again for the valuable time and constructive feedback provided during this review.

We sincerely thank Reviewer GV1w for the constructive feedback and for recognizing our contributions in introducing a new benchmark with structured attribute annotations, an interpretable model, and comprehensive evaluations. Below, we respond to your comments and suggestions in detail.

---

Q1: "Suggestion on the structured formulation of referring expressions and whether it limits generalization to more comprehensive or free-form sentences."

A: We appreciate your observation. Our choice to structure the referring expressions using four attributes — Appearance, Status, Relation-to-Viewer, and Relation-to-Others — was driven by two practical considerations:

1. To enable interpretable, fine-grained supervision of multimodal grounding;
2. To support real-world referential behaviors, where such attributes commonly appear in driving scenarios (e.g., “the pedestrian crossing the street on the left”).

The expressions in Talk2Event are not rigid templates. They are generated using Qwen2-VL under a context-rich prompting scheme, and then filtered and validated by humans to ensure fluency and diversity. As shown in the word clouds (see Fig. 2 in the manuscript) and examples in the Appendix (see Section A.4 in the supplementary material), expressions naturally vary in phrasing and length, even while preserving attribute compositionality.

In short, while our formulation encourages attribute-rich descriptions, the actual expressions exhibit broad linguistic variation, making them suitable for generalization beyond strict structural forms. As suggested, we will continue to enhance expression diversity in future extensions.

---

Q2: "Concern that the model is too structurally dependent (on four experts) and may not generalize well to more complex or open-ended language inputs."

A: Thank you for this insightful point. The four-expert design of EventRefer is motivated by the four grounding attributes annotated in Talk2Event, enabling the model to explicitly reason over appearance, motion, and relational cues — aspects that are especially important in real-world, dynamic scenes.

While this design introduces attribute-specific structure, it does not enforce hard constraints on the input format. In fact:

• The fuzzy matching mechanism in Section 4.1 is soft and robust to paraphrasing, allowing the model to handle a wide range of expressions.
• The MoEE module adaptively weighs attributes based on their informativeness per sample, regardless of sentence complexity.
• The decoder remains unified, producing a single prediction per sample without branching logic.

Thus, the model is attribute-aware but not attribute-restricted, and generalizes well to free-form language input. As suggested, we have refined the explanation in Section 4 to clarify this design choice.

---

Q3: "Model is too complicated."

A: We appreciate this concern. While EventRefer incorporates multiple components, such as positive word matching, attribute-aware masking, and mixture-of-experts fusion, the design is intentionally modular and lightweight:

• Each module builds on standard components (e.g., fuzzy matching, Transformers), with minimal architectural overhead.
• The MoEE module introduces few additional parameters and shares backbone features across experts, avoiding redundancy.
• During both training and inference, the decoder remains compact, generating one box per sample without structural branching.

This design balances interpretability with efficiency, enabling compositional reasoning while keeping the system tractable. We acknowledge the potential for simplification and plan to explore more compact or LLM-driven architectures in future work.

---

Q4: " Event-based LLMs like EventGPT are a more general way of approaching the problem. As they are more capable of freely interacting with a chat-based framework."

A: Thank you for raising this broader perspective. We agree that EventGPT represents a promising and general direction for multimodal, conversational reasoning. Our work, however, addresses a complementary challenge: precise, interpretable referring expression grounding, with a strong focus on spatial and attribute-level localization. This design choice enables transparent interpretation and targeted ablation, which is challenging with end-to-end black-box LLMs.

Moreover, current LLMs typically operate within language-generation settings and lack the architectural supervision required for precise box prediction. Existing models, e.g., EventGPT, have to call the corresponding tools for completing downstream tasks (such as calling Grounding-DINO for detection and Grounded-SAM for segmentation). Instead, our formulation serves as a complementary line of this research, enabling spatially grounded understanding with interpretable components, which could later be integrated with general-purpose LLMs for more flexible interaction.

As suggested, we have properly discussed these aspects in our manuscript (in both Section 1 and Section 2) and acknowledged the own advantage of EventGPT and our EventRefer approach.

---

Q5: "No comparison with EventGPT."

A: We appreciate the suggestion. However, as of this submission, EventGPT has not been designed for spatially grounded tasks such as referring object grounding or box prediction. It focuses on language-only outputs, e.g., question answering, captioning, or dialogue. As mentioned in Q4, existing models, e.g., EventGPT, have to call the corresponding tools for completing downstream tasks (such as calling Grounding-DINO for detection and Grounded-SAM for segmentation).

Our task, by contrast, is referring expression grounding with bounding box outputs, requiring spatial supervision and localization performance, for which no existing event-based LLMs currently provide support.

In future work, we plan to explore hybrid approaches that combine EventRefer with instruction-tuned VLMs for grounded reasoning.

We agree with you that the comparison is valuable. As suggested, we have added a discussion in the revised Related Work section (Section 2), outlining the complementary nature of EventGPT-style approaches and our attribute-aware grounding framework.

---

Q6: "Evaluation is only performed on Talk2Event, which follows a structure similar to the training data."

A: Thank you for pointing this out. We acknowledge that our current evaluation is conducted on the Talk2Event benchmark, which is indeed designed using the same attribute-based supervision as the training data. This in-distribution evaluation format is consistent with established practice in visual grounding, as adopted in prior works such as RefCOCO, ScanRefer, and M3DRefer — where models are trained and tested on splits drawn from the same dataset.

In addition:

• Talk2Event was curated with explicit diversity goals, including variation in lighting, motion, object classes, and expression phrasing.
• Our experimental protocol includes modality shifts (event-only, frame-only, and fusion), providing strong generalization signals even within the benchmark.
• As shown in our attribute activation and ablation study, the model responds differently depending on the scene context — demonstrating adaptive generalization, not overfitting to a fixed structure.

We agree that broader cross-benchmark evaluation (e.g., testing on expressions from unrelated datasets or VLM-generated instructions) would offer further insights into out-of-distribution robustness. As suggested, we have added this discussion as a future direction in the revised manuscript (Section 2) and supplementary material (Section D).

---

Q7: "How is the event encoded? What is the duration of the event context?"

A: Thank you for the question. Following standard event-based perception practices (e.g., RVT and DAGr), we discretize the asynchronous event stream into a spatiotemporal voxel grid over a short temporal window. Specifically:

• The voxel grid includes $T = 10$ temporal bins, covering a time span of $400$ milliseconds centered at timestamp $t_0$.
• Each voxel encodes polarity ($\pm 1$), spatial position, and time bin index, resulting in an input $E ∈ ℝ^{2×T×H×W}$.

This $400$ ms window provides sufficient context for capturing short-term dynamics while maintaining low latency for real-time applications. As suggested, we have made this clearer in Section 3.1 in the manuscript and Section B.2 and Section B.4 in the supplementary material.

---

References:

• [EventGPT] Liu, S., Li, J., Zhao, G., Zhang, Y., Meng, X., Yu, F. R., ... & Li, M. (2025). "EventGPT: Event Stream Understanding with Multimodal Large Language Models." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 29139-29149).
• [RVT] Gehrig, M., & Scaramuzza, D. (2023). "Recurrent Vision Transformers for Object Detection with Event Cameras." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 13884-13893).
• [DAGr] Gehrig, D., & Scaramuzza, D. (2024). "Low-Latency Automotive Vision with Event Cameras." Nature, 629(8014), 1034-1040.

---

Last but not least, we would like to sincerely thank Reviewer GV1w again for the valuable time and constructive feedback provided during this review.

We sincerely thank Reviewer ZmP8 for devoting time to this review and providing constructive feedback. We are encouraged that you recognize Talk2Event as a timely benchmark with a well-designed dataset and interpretable model. Below, we address your comments in detail.

---

Q1: "Potential noise introduced by color descriptions in expressions when using event-only data."

A: Thank you for mentioning this point. Indeed, event cameras do not capture color information, which might suggest a potential mismatch with language expressions that include color attributes. However, we intentionally retain such descriptions in Talk2Event for several reasons:

• Modality generality: Talk2Event supports three grounding configurations — event-only, frame-only, and event-frame fusion. Color attributes remain crucial for the latter two.
• Robust modeling: Our EventRefer design allows the model to learn to down-weight uninformative cues (e.g., color in event-only mode) via the MoEE module, which adaptively selects the most relevant attribute experts.
• Realistic deployment: In real-world robotics, systems often integrate multiple sensors. Retaining full language fidelity encourages the development of robust, cross-modal models that gracefully degrade when specific cues are absent.

Nevertheless, we fully agree with you that a variant without color descriptions could be valuable for event-only research. Following your suggestion, we are currently exploring releasing such filtered expressions as an additional modality-specific split to further support this use case.

---

Q2: "Clarification on symbol overload of '$C$' (L153 vs. L161) and how the element-wise product in L165 operates across mismatched shapes."

A: Thank you for pointing out the overloaded use of the symbol $C$. We have corrected this in the revised version by changing the token sequence length symbol from $C$ to $T_\mathrm{len}$ (i.e., $T_\mathrm{len}$ for token length, and $C$ remains as the channel dimension). This resolves the ambiguity and improves readability.

Regarding the element-wise product in Line 165: although we use shorthand notation $H^\mathrm{att}_i = m^\mathrm{att}_i ⊙ H$, the actual masking involves aligning token-level attention priors to query-level hidden states. Specifically:

• The token-level masks $m_i$ and $m_0$ are used to guide attention or generate soft scores.
• These are then aggregated (e.g., via attention pooling or alignment with text-query cross-attention) into query-level weights, forming $m^\mathrm{att}_i ∈ ℝ^{B × Q × 1}$.
• This mask is then broadcasted across the channel dimension and applied to $H ∈ ℝ^{B × Q × C}$ to produce $H^\mathrm{att}_i$.

As suggested, we have clarified this pipeline and provided additional details in the revised text and supplementary material to ensure good reproducibility.

---

Q3: "Clarification on how the query-level mask $m^\mathrm{att}_i$ is derived from the token-level map $m_i$, and the shapes of $m_i$, $m_0$, and $m^\mathrm{att}_i$."

A: Thank you for highlighting this important question. Here is a clarification of the shapes and derivation:

• $m_i ∈ ℝ^{B×T_\mathrm{len}}$: soft token map for attribute $δ_i$ (computed via fuzzy matching over token indices).
• $m_0 ∈ ℝ^{B×T_\mathrm{len}}$: token mask for public (non-attribute) context.
• These are combined as $m_i ∨ m_0$ to form a textual attention prior, used to guide attribute-specific reasoning in the text modality.

To apply attribute masking on the fused query-level hidden states $H ∈ ℝ^{B×Q×C}$, we construct $m^\mathrm{att}_i ∈ ℝ^{B×Q×1}$ — a query-level soft mask.

While the manuscript uses simplified notation, this query mask is not directly derived from $m_i$, but rather formed implicitly from downstream fusion between event queries and language features (informed by $m_i$). Specifically, the language encoder embeds the attended tokens (weighted by $m_i$), and the cross-modal encoder maps this into the query space. We apply $m^\mathrm{att}_i$ as a gating mask over $H$ to retain attribute-relevant query features.

As suggested, we have clarified this process in Section 4.1 and included an expanded explanation and figure in the supplementary material for better clarity.

---

Q4: "Suggestion for fine-tuned results of frame-only baselines (e.g., OWL, GroundingDINO, YOLO-World) for fairer comparisons."

A: We fully agree that a fine-tuned evaluation of frame-only baselines would provide a more balanced comparison. In our current submission, we prioritized the zero-shot results to assess the generalization of large open-vocabulary models, as prior work (e.g., OWL-ViT) emphasized this setting.

We are currently conducting the fine-tuning experiments for GroundingDINO and OWL-ViT on Talk2Event, and will include these updated results in an updated version of Table 2. We note, however, that these models are typically pre-trained on large-scale static image datasets (e.g., COCO, LVIS, and proprietary collections), which differ substantially from our dynamic, driving-centric event-frame grounding setup. This domain gap might pose challenges for fair comparisons, even after fine-tuning.

Nevertheless, we recognize that including such comparisons will make the evaluation more comprehensive. Therefore, we have carefully documented training protocols, domain adaptation challenges, and any observed failure cases in the revised supplementary file, to provide an assessment of how generalist models perform under our event-based grounding scenario.

Based on your suggestion, we commit to updating the evaluation accordingly to further strengthen the completeness and transparency of our benchmark study.

---

Last but not least, we would like to sincerely thank Reviewer ZmP8 again for the valuable time and constructive feedback provided during this review.