acceptodds
Under review as a conference paper at ICLR 2027

ForeWarn: Risk-Adaptive Inference for Proactive Safety Warning in Streaming Video

11 views

Abstract

Streaming video models now describe what is happening in front of a camera, and proactive safety warning asks them to act on what is about to happen. However, existing warning methods decide from the current state, and confirming a hazard once it is visible consumes part of the time left to warn. Our diagnostics reveal that serial confirmation filters out early candidates and that a later confirmation recognizes more but returns too late to warn. In this paper, we propose ForeWarn, which lets the near-future risk predicted by a small vision-language model decide both when to warn and when to verify, so that a forecast acts before confirmation where time is short and waits for it where time allows. The same risk read at two horizons drives an immediate path that does not wait for a verifier and an advisory path that combines predicted risk with a returned state judgment; a risk-dependent schedule decides when to query the verifier. Furthermore, calibration on normal footage sets the advisory thresholds per deployment, and a one-step analysis ties the value of verification to confirmation quality and timely return. Extensive experiments on five public domains, self-recorded live trials, and three Jetson edge devices demonstrate that ForeWarn raises macro pre-incident recall to 1.4 that of the strongest of four recent baselines at similar verifier cost, with gains of up to 47.7 points on a single domain and up to 110 less verifier time per clip.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
proactive safety warning, streaming video understanding, vision-language models, risk-adaptive inference, edge deployment

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.80 over 5 reviews, above 85% of the others.
Reviewer_h1sVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies proactive safety warning in streaming video, where systems generally need to issue warnings before an incident actually occurs. The authors propose ForeWarn, which uses a small VLM to predict near-term incident risk and uses the risk score to determine both whether to issue a warning immediately and when to call a verifier. When short-horizon risk is high, the system issues a warning without waiting for the verifier. For longer-horizon risks, the verifier can confirm whether a hazardous state is present. The authors report that, at a comparable false-alarm rate, ForeWarn achieves higher pre-incident recall while reducing verifier calls; however, its generalizability across scenarios remains unestablished.

Critical strengths

The problem and the proposed design are well aligned. The paper distinguishes between a verifier’s accuracy and whether its results arrive in time, arguing that online warning depends not only on correct recognition but also on delivering a warning while action is still possible. Short-horizon risk triggers an immediate warning, while longer-horizon risk is combined with verification results to issue an advisory warning, addressing both fast-onset incidents and slowly developing hazards.

Critical weaknesses
  1. The key risk scores are not probability-calibrated. The paper states that r3 and r5 are used as scores rather than calibrated probabilities. Thus, r3=0.9 does not mean that the true probability of an incident is 90%. More extensive experiments are needed to establish how stable the thresholds are across scenarios and to justify how they are selected.
  2. The problem is valuable, but the distinction from prior work needs to be clarified. Fast–slow models, selective model invocation, and inference cascades have been studied in related work. The authors should explain what their design adds beyond existing fast–slow pipelines and selective VLM routing, and why these mechanisms cannot be reduced to generic confidence-based routing.
Questions for authors
  1. How was the immediate-warning threshold Θu=0.9 chosen? Was it selected using only the development set? Did the authors evaluate recall, false-alarm rate, and balanced accuracy across different threshold values?
  2. Given that r3 and r5 are not probability-calibrated, how stable are the thresholds across datasets or cameras? Did the authors validate the method in independent scenarios beyond those used for deployment-specific calibration?
  3. Could the authors report confidence intervals for the main aggregate results and quantify how end-to-end device latency affects timely recall?
5 comments
AuthorsOpenReview ↗

Response to Reviewer h1sV (1/2)

Thank you for the review and for raising the ethics point in a constructive way. References below are to the revised paper, where changes are in blue. R / F denotes pre-incident recall and the clip false-alert rate (FAR) on held-out normal clips, and α is the nominal clip-FAR budget of site calibration.

We first note one fact about the summary. The macro recall of 0.670 against 0.473 for StreamSense comes at a macro clip FAR of 0.362 against 0.291, with similar verifier calls. With StreamSense re-thresholded to at most ForeWarn's clip FAR, the lead is 13.3 points, and it comes from the staged falls and EDF (Section 4.2, Table 10). Two StreamSense rows in the submission had used the zero-shot scout, and the revision corrects them (Table 1).

W1. The risk scores are not calibrated probabilities, and threshold stability and selection need more evidence.

The thresholds are operating points and not calibrated probabilities. The revision reports the reliability of the scores and adds a rule that sets the immediate threshold from calibration normals (Appendices H.8 and H.2).

  • Not probabilities. The expected calibration error of the 3 s risk score is 0.145 to 0.279 across domains, so the policy uses the score for ranking. Its AUROC ranges from 0.858 on the falls to 0.568 on OCCU.
  • Stability. On the falls and EDF, 0.9 is the first grid value at which no held-out normal triggers the immediate path. It corresponds to a quantile of normal scores between 0.968 and 1.000 that differs in each domain, so no single quantile reproduces it (Appendix H.2).
  • What we change. The new rule sets the immediate threshold from calibration normals, like the advisory thresholds. The main results keep the frozen 0.9, and the operating points of the new rule are reported next to them.

W2. What does the design add beyond fast-slow pipelines, selective invocation and cascades, and why is it not generic confidence routing?

ForeWarn differs from a cascade in its signal and in its action, and the difference is measurable. With the same scout, verifier and calibration rule, a controller in which every alert waits for the verifier loses 7 of 31 clips on the staged falls and 19 of 42 on EDF (Section 4.3, Appendix H.4). In a cascade or a confidence router, the small model's uncertainty decides whether the large model is consulted, and the output on that step waits for the large model. ForeWarn differs in two ways. First, its signal is a forecast of time-to-incident at two horizons and not confidence in the present answer. Second, the path with the least time never waits. The short horizon issues the warning from the scout, and verification runs asynchronously where it can only add an advisory. On the other three domains the two controllers are close. At tight budgets the second model adds recall on long-runway footage. On the long-runway home pool Home-L at α = 0.3 the full policy gives 0.370 / 0.432, against 0.204 / 0.318 for a single threshold on the scout (Appendix H.3). Section 2 now positions the method against cascades and selective routing in these terms.

Q1. How was θu = 0.9 chosen?

It was fixed before the benchmark runs from a sweep on EDF recordings, and the other four domains inherit the value without tuning. Those recordings are also the EDF evaluation videos, so the EDF results carry selection bias. The revision states this in Section 3.3 and reports a new sweep of the threshold on all five domains (Appendix H.2). On the falls the new sweep gives 0.968 / 0.312 at a threshold of 0.80, 0.935 / 0.125 at 0.90 and 0.871 / 0.125 at 0.95. On the falls and EDF a higher threshold only removes recall, and a lower one adds a little recall at a much higher false-warning rate.

Q2. Stability across datasets or cameras, and validation beyond the scenarios used for calibration.

The advisory thresholds are site-specific by design, and one calibration pass on normal footage sets them. An additional set of 259 clips adds two domains outside the scout's training sources (Appendices H.8 and H.10).

  • Across datasets. A gate carried over from another domain gives a clip FAR of 0.838 to 0.973 on OOPS and 0.466 to 0.636 on Home. Each site is therefore calibrated on its own normal footage.
  • Across cameras. Calibration and evaluation use disjoint normal clips within a domain. The public sets give no camera-level split, so stability across cameras remains to be tested (Appendix F.3).
  • Additional set. With thresholds calibrated on its own normal clips, the set of 259 clips gives 0.309 / 0.282 and 0.414 / 0.436 at α = 0.2 and 0.3. Its industry and health groups are outside the scout's training sources, and its other groups share sources with our data. The controlled recordings are a separate site, and their result is given under Q3.
AuthorsOpenReview ↗

Response to Reviewer h1sV (2/2)

Q3. Confidence intervals for the main results, and the effect of device latency on timely recall.

Wilson intervals for every domain are in Appendix H.6, and the device experiments are in Appendix H.9. The macro recall difference against StreamSense is +0.197, with an interval from +0.138 to +0.257. The clip-FAR intervals are wide because the normal pools hold 16 to 88 clips. In the recorded device runs accumulated lag roughly halves the timely hits on every tier, from 117 to 56 of 299 on the 64 GB tier. A new control experiment on that tier has an on-board baseline of 70 timely hits. Running the verifier off the board on a separate server raises them to 176, with 0.519 of normals alert-free against 0.803. The controlled recordings stay at 1 to 2 of 35 on the two larger tiers, and Section 4.5 reports these as decision misses and not lag.

Ethics flag: consent, privacy and review for the self-recorded trials.

The submission did not describe these three points, and we answer them in turn.

  • Who was recorded, and consent. The 65 controlled recordings were staged and recorded by members of our laboratory. Everyone who appears is a researcher on this project or a student of the same laboratory. No one outside the laboratory was recruited, and no member of the public was filmed. Each person who appears agreed orally to the use of the recordings in this study.
  • Privacy. The recordings are used only for the execution evaluation of Section 4.5. They are kept on our group's storage and are not part of any released dataset. No frame of them appears in the paper. The release holds only their per-step scores, decisions and timing, together with a manifest of clip codes, labels, incident times and durations. It holds no video, image or audio.
  • Ethics review. The recordings were not submitted for institutional ethics review. The participants were limited to the researchers on this project and students of our laboratory, as stated above. If the ethics reviewers consider a formal determination necessary, we will request one from our institution and report its outcome.

Our Ethics Statement described the recordings as "volunteer enactments" and noted that the archive holds no consent or institutional-review documentation. That wording did not say who the volunteers were or whether a review took place, and we have rewritten the statement to give the points above directly.

AuthorsOpenReview ↗

Follow-up to Reviewer h1sV

Dear Reviewer h1sV,

Thank you again for your review. Our response is posted above, together with a revised PDF. The revision reports how the immediate threshold was chosen and adds a sweep over it on all five domains (Section 3.3, Appendix H.2). It positions the method against cascades and confidence routing with a direct comparison (Sections 2 and 4.3), and it adds intervals for the main results (Appendix H.6). The Ethics Statement now describes who was recorded, consent and review status.

Could you let us know whether these changes address your concerns, including the ethics point, or which points remain open? We will answer any further question before the discussion closes on November 18. If the concerns are resolved, we would be grateful if you could reflect this in your assessment.

Best regards,

The authors

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

This paper addresses proactive safety warning for streaming video, whose goal is to issue alerts before physical incidents happen rather than only detecting visible hazards. The authors first diagnose two critical failure modes of existing verification‑gated pipelines: serial verification filters many early forecast candidates, and late verification yields accurate judgements but returns too late for intervention. They propose ForeWarn, a risk‑adaptive inference framework driven by two‑horizon risk scores from a lightweight vision‑language scout model. ForeWarn maintains two warning paths: an immediate urgent path that skips heavy‑weight verifier confirmation for imminent hazards, and an advisory path gated by asynchronous verifier state together with long‑horizon predicted risk. It employs a risk‑dependent schedule to trigger costly verifier queries adaptively instead of fixed intervals. A per‑deployment calibration scheme tunes advisory thresholds purely on normal hazard‑free footage without accident labels, alongside a theoretical one‑step utility analysis to judge when verification is worth its latency cost. Evaluations cover five public video domains, self‑recorded live trials and three Jetson edge hardware tiers. Compared against four recent baselines, ForeWarn achieves 1.4× macro pre‑incident recall under comparable verifier computation overhead; ablation validates the necessity of dual warning paths and adaptive scheduling. Real‑device measurements reveal accumulated runtime lag substantially degrades timely warning performance even when offline replay looks promising.

Critical strengths

The paper quantitatively diagnoses that mandatory heavy‑model verification for all warnings discards early forecast candidates and yields belated high‑confidence confirmations missing intervention windows, explaining pre‑incident recall shortcomings of prior safety systems largely overlooked by accuracy‑centric benchmarks, and it addresses this temporal‑latency dilemma via a practical dual‑path risk‑adaptive inference design: imminent hazards fire non‑revocable immediate alerts while slowly‑building risks undergo asynchronous verifier checks, paired with risk‑driven verifier scheduling, normal‑footage‑only deployment calibration and a theoretical utility criterion to assess costly verification under tight temporal deadlines.

Critical weaknesses
  1. ForeWarn improves pre‑incident recall at the cost of elevated clip‑level false‑alert rate (FAR). On the OCCU occluded domain, overall balanced accuracy drops below 0.5. The scout model’s performance severely degrades under visual occlusion, and neither the scout nor verifier can reliably generate risk signals when key visual cues are blocked. The paper acknowledges this failure case but provides no mitigation strategy. This weakness matters because real‑world monitoring frequently suffers from partial occlusion, which limits real‑world practical reliability.
  2. While offline replay yields strong benchmark numbers, shared‑hardware resource contention on Jetson edge devices creates severe accumulated lag, roughly halving the achievable timely hits. The paper only identifies this gap but does not propose scheduling optimizations to mitigate throughput‑induced lag for end‑to‑end deployment. Since streaming safety warning is latency‑critical, high replay performance cannot translate into usable real‑world performance without addressing runtime resource competition, which weakens the practical contribution for edge deployment scenarios.
Questions for authors

See Weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer Mnop

Thank you for the review. We tested mitigations for both weaknesses and report them below. References are to the revised paper, where changes are in blue. R / F denotes pre-incident recall and the clip false-alert rate (FAR) on held-out normal clips, and α is the nominal clip-FAR budget of site calibration.

W1. OCCU: balanced accuracy below 0.5 and no mitigation.

OCCU is hard for every method. The revision explains why and reports three mitigations that we tested (Section 4.2, Appendices E.5, H.2 and H.4).

  • What makes OCCU different. OCCU is the occluded counterpart of EDF, and the end of each fall is hidden behind an occluder. The four baselines recall at most 0.048 on it, and ForeWarn recalls 0.119. On the staged falls and EDF, where the fall is in view, the same policy recalls 0.935 and 0.929.
  • Diagnosis. The scout ranks occluded falls poorly, with an AUROC of 0.568 on OCCU against 0.858 on the staged falls (Appendix H.8). The verifier withholds confirmation on views that it cannot read (Section 4.3).
  • Mitigations. Lowering the immediate threshold to 0.8 raises OCCU recall from 0.119 to 0.548, at a clip FAR of 0.524 against 0.286. A scout-uncertainty gate and a longer verifier window do not help (Appendix H.4).
  • Outlook. Used as the scout with a single threshold, a vision-only probe recalls 0.595 to 0.690 on OCCU at α = 0.1. This suggests that occluded falls can be detected from the footage, although it is one probe on a small evaluation set.

W2. Contention on shared edge hardware halves timely hits, and there is no scheduling optimization.

We evaluated three scheduling options on the device. Running the verifier off the board on a separate server raises timely hits from 70 to 176 of 299 on the 64 GB tier. The share of alert-free normals falls from 0.803 to 0.519 (Appendix H.9).

  • Scheduling options. Off-board verification gives 176 timely hits, skipping stale steps 87, a 1 s scout step 75 to 77, and all three together 144. With the verifier off the board no lag accumulates, so the longer step and the smaller budget only reduce the number of decisions. On matched queries the on-board and off-board verifiers give the same answers, with an agreement of 0.994. The gain therefore comes from answers that arrive in time. With off-board verification 96% of queries are answered before being superseded, against 27% on the board.
  • Source of the lag. Contention between the two models on one board explains a minor part of the recorded lag, and a rerun on the recorded inputs does not reproduce the rest (Appendix H.9).
  • Requirement. Short-runway warning needs a scout step faster than the available lead of about 1 s (Section 4.5). On the two larger tiers the controlled recordings stay at 1 to 2 of 35, because those misses are decision misses, which scheduling does not address.
AuthorsOpenReview ↗

Follow-up to Reviewer Mnop

Dear Reviewer Mnop,

Thank you again for your review. Our response is posted above, together with a revised PDF. For OCCU, the revision explains what makes the domain hard and reports the mitigations that we tested (Section 4.2, Appendices E.5, H.2 and H.4). For execution on shared hardware, it adds device experiments with three scheduling options (Appendix H.9).

Could you let us know whether these changes address your two concerns, or which points remain open? We will answer any further question before the discussion closes on November 18. If the concerns are resolved, we would be grateful if you could reflect this in your assessment.

Best regards,

The authors

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Reviewer_bA3rOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes “ForeWarn,” an active early warning system designed for streaming video. The authors think detection and verification are two critical processes in an early warning system, but both are time-consuming, which is unacceptable in some urgent situations. Therefore, the authors designed a framework consisting of a 2B VLM and an 8B VLM. When the small model detects an imminent danger, "the short-term risk is high", it issues an alert immediately. Only when the risk spans a longer time horizon does the system invoke the large model to perform iterative verification and determine whether an alert is necessary. This approach is somewhat similar to a “fast-slow” system design with route selection. Across the five dataset scenarios collected by the authors, ForeWarn outperforms the current state-of-the-art methods and improves operational efficiency in some cases. (0.478 to 0.670) The paper also reports substantial false alarms and performance degradation caused by accumulated processing lag on edge devices. I could not locate a code repository or code supplimentary in the materials I reviewed.

Critical strengths
  1. This paper addresses a practical issue: many existing danger-warning systems for daily life fail to adequately account for the time required for detection and verification processes.
  2. Despite some issues with the comparison methodology, it is evident that there are improvements in areas such as recall and hazard identification.
  3. The experiments cover public benchmarks, custom-recorded real-time trials, and three different edge device configurations; this is good, as it allows for a multi-faceted evaluation.
  4. The system proposed in this paper has a real-world implementation.
Critical weaknesses
  1. The paper does not seem to sufficiently justify the necessity of a dual-system architecture. The experimental section lacks comparisons against a single "small" VLM system (utilizing the same dual-stream data flow) or a single "large" VLM system (using the same data flow).
  2. It is unclear what the specific incremental methodological contribution of this work is compared to StreamSense. While the results show superior performance, the innovations appear to be minor, detail-oriented, or primarily engineering-focused. What are the core methodological innovations? I would appreciate it if the authors could clarify this briefly.
  3. There are many methods for assessing short-term risks. In this paper, the authors rely on a 2B-parameter VLM, which makes the role of the language model seem somewhat superfluous. Have the authors considered using alternative visual representation models,instead of a VLM, to determine whether a scene should be passed to a larger VLM for reasoning?
  4. Most of the innovations in this paper center on the collaborative design of small and large models, rather than algorithmic improvements in representation extraction or predictive modeling. I am not convinced that this focus aligns well with the core themes of ICLR. The authors should address this point when clarifying their methodological contributions.
  5. While ForeWarn improves recall, it also leads to an increase in false positives. Furthermore, the evaluation should include a comparison where the number of alerts generated by ForeWarn is adjusted to match that of StreamSense. Simply increasing the false-positive rate of StreamSense only demonstrates an advantage at higher false-positive levels; it fails to prove that ForeWarn performs better at the lower false-positive levels established by the baseline.
  6. The test dataset used in this study is quite small; it is uncertain whether the observed gains stem from actual improvements or merely from noise/perturbations, or perhaps because the model has learned unrelated patterns.
  7. Could the authors provide the "early warning time" (lead time) for the different models? I believe this is a crucial metric.Please provide detailed results (like the time-to-danger output in the same case that both model detect)
  8. The results from the real-world deployment appear underwhelming; on the R65 dataset, both larger-scale configurations successfully detected only 1 out of 35 hazardous events in a timely manner. While the author’s effort to implement a deployment is commendable, the current outcome does not appear to constitute a meaningful contribution, as there is no design consideration for GPU scheduling or practical, real-world usage. Furthermore, a rate of 95 false positives per hour would be excessively disruptive to users. Could the author refine the system to keep the number of false positives per hour within a reasonable range before presenting comparisons of model performance? (For instance, I would like to see the performance results when the number of false positives is limited to, say, fewer than 12 per hour.)
Questions for authors

Some of these points have already been mentioned in the "weaknesses" section; I am reiterating them here. I suggest that the authors address these specific questions in their rebuttal.

  1. Please briefly explain the incremental improvements of your approach compared to the core methodology of StreamSense.

  2. Under identical control logic and false-positive constraints, is it possible to verify the additional benefits of the dual-VLM architecture and the language modeling component by comparing it against baselines using a single small VLM, a single large VLM, and a standalone visual representation model?

  3. Does ForeWarn still maintain a performance advantage once its false-positive rate is reduced to the native level of StreamSense? Please also provide a comparison under stricter false-positive budgets—for instance, a limit of no more than 12 false positives per hour.

  4. Please report the lead time for early warnings and the actual timely recall rate for comparable methods. Specifically, we would like to see paired results for events detected by both systems, including the timestamp of the first actionable warning, the actual onset of the hazard, and the resulting lead time. Additionally, please clarify whether there are any operational configurations that can effectively mitigate processing backlogs on edge devices.

  5. Given the small scale of the test, could you provide paired confidence intervals for key performance differences, along with an analysis broken down by data source and scenario, to demonstrate the stability and generalizability of the results?

  6. Regarding the baseline comparison: When implementing the StreamSense baseline for your experiments, did you adapt it specifically for hazard-warning data, or did you primarily port over its routing mechanism? Furthermore, was the model trained for 10 epochs, as in the original StreamSense implementation? What dataset was used for this training?

8 comments
AuthorsOpenReview ↗

Response to Reviewer bA3r (1/3)

Thank you for the detailed review and the numbered questions. We answer each weakness together with the question that repeats it. References below are to the revised paper, where changes are in blue. R / F denotes pre-incident recall and the clip false-alert rate (FAR) on held-out normal clips, and α is the nominal clip-FAR budget of site calibration.

We first note two facts about the summary. All benchmark videos come from public datasets: OmniFall, EDF, OCCU, OOPS and SmartHome-Bench. Only the 65 controlled recordings of Section 4.5 are ours. The macro recall of 0.670 against 0.473 for StreamSense comes at a clip FAR of 0.362 against 0.291, with similar verifier calls. With StreamSense re-thresholded to at most ForeWarn's clip FAR, the lead is 13.3 points, and it comes from the staged falls and EDF (Section 4.2).

W1 and Q2. Is the dual-VLM design necessary? Single small VLM, single large VLM, vision-only model under the same control logic and false-positive constraints.

We ran all three arms, and Section 4.3 of the revision reports them. The second model adds recall on long-runway footage at tight budgets, and at α = 0.5 it lowers the clip FAR on the falls and EDF. A single large VLM is too slow for short-runway events, and the fine-tuned small VLM ranks better than a vision-only probe at low false-alarm rates.

  • Single small VLM, no verifier (Appendix H.3). At budgets of 0.2 and above it reaches ForeWarn's recall on the falls, EDF and OOPS, and it recalls more on OCCU. Its clip FAR on the falls and EDF is higher at α = 0.5 and similar at α = 0.2, with 0.968 / 0.562 against 0.935 / 0.125 on the falls at α = 0.5. On the long-runway home pool Home-L the pair recalls more at tight budgets, with 0.370 / 0.432 against 0.204 / 0.318 at α = 0.3.
  • Single large VLM (Appendix H.3). Polled every second and charged its recorded latency, the frozen 8B recalls 0.129 on the falls and 0.000 on EDF. Charged 0.483 s per call it recalls 0.560 on Home against 0.486 for ForeWarn, with 26.5 calls per normal clip against 7.51.
  • Vision-only scout (Appendix H.4). A V-JEPA 2 ViT-L probe can take the scout's role at α = 0.5 once the immediate threshold is calibrated for it. At α = 0.1 the fine-tuned 2B recalls more, with 0.935 against at most 0.355 on the falls. The probe was trained on a different corpus, so the comparison is not controlled for modality.

W3. Is the language model in the 2B scout superfluous? Could a visual representation model decide what to pass to the larger VLM?

A visual representation model can fill the scout's role, and Appendix H.4 reports one. What the fine-tuned VLM adds is ranking quality at low false-alarm rates. The policy reads only five time-to-incident class probabilities from its language head. Its window-level AUROC is higher than the vision-only probe's in four of five domains, including 0.862 against 0.787 on the falls. Its recall at α = 0.1 is higher on the falls, EDF and OOPS (Appendix H.4). The probe is faster, at 0.054 s per window against 0.191 s, and it recalls more on OCCU.

W2 and Q1. What is the methodological increment over StreamSense?

There are three changes to the decision rule, and the second carries the measured gain.

  1. The control signal is a forecast. StreamSense routes each step by the small model's confidence in its current answer. ForeWarn reads predicted time-to-incident at two horizons.
  2. The short-horizon warning never waits for the large model. ForeWarn warns from the scout on the short horizon, and verification runs asynchronously where it can only add an advisory. With identical models and calibration, a controller in which every alert waits for the verifier recalls 0.645 against 0.871 on the falls and 0.429 against 0.881 on EDF (Appendix H.4).
  3. Verification is scheduled by long-horizon risk and the time since the last query, and the operating point is calibrated per site against a total false-alert budget.

Against StreamSense with the same scout and verifier, ForeWarn leads on the falls and EDF at matched false-alarm levels (Table 10).

Q6. How was the StreamSense baseline implemented?

We ported its routing rule unchanged. We run the authors' released routing script on the same fine-tuned scout and the same 8B verifier as ForeWarn. The router has no trainable parameters, so no training was involved. At each step it accepts the small model's prediction, invokes the large model, or defers. It decides with two thresholds that are closed-form functions of the time since the last accepted decision. Its two hyperparameters keep the released defaults of 6 and 3.0, and deferred reads are included in its call counts. Two of its rows in the submission had used the zero-shot 2B as the small model. The revision corrects them, which moves its OOPS row to 0.907 / 0.784 and its macro recall to 0.473 (Table 1, Appendix D.2).

AuthorsOpenReview ↗

Response to Reviewer bA3r (2/3)

W4. Fit with ICLR.

The paper studies a decision problem that arises once a learned forecaster exists: when should a forecast act without confirmation, and when is a second model's judgment worth its delay? Its contributions are the two-path decision rule, a one-step criterion for when verification can help, and an evaluation protocol that charges inference latency against recall. Inference-time decision rules for learned models and their use in a safety task are within the scope of the conference. The introduction of the revision states the contribution in these terms.

W5 and Q3. Does ForeWarn still lead once its false positives are reduced to StreamSense's native level? A budget of at most 12 per hour.

Yes on the staged falls and EDF, which carry the lead. On OOPS and Home the two methods are close (Appendices E.1 and H.1).

  • Falls. At α = 0.1 ForeWarn's held-out point is 0.903 / 0.062, the same clip FAR as StreamSense's native 0.645 / 0.062.
  • EDF. StreamSense's native point is 0.452 / 0.095, and ForeWarn's lowest-FAR calibrated point is 0.690 / 0.000.
  • OOPS. StreamSense's native clip FAR of 0.784 is above ForeWarn's 0.676. Re-thresholded to at most ForeWarn's FAR it recalls 0.867 against 0.880.
  • Home. ForeWarn's calibrated point at α = 0.4 is 0.349 / 0.432, against StreamSense's native 0.312 / 0.466. With StreamSense re-thresholded to ForeWarn's clip FAR at each budget, it recalls more at four of the five budgets, by at most five clips.
  • OCCU. Both methods recall little under occlusion.
  • At most 12 per hour (Appendix H.1). On the falls and EDF this budget means zero false warnings. With each method's setting chosen on the test normals, ForeWarn keeps 0.871 and 0.905 there, against 0.387 and 0.381 for StreamSense. With its operating point chosen on calibration normals only, ForeWarn gives 0.690 on EDF with no false warning and 0.839 on the falls with one false warning step, which is above the budget. On Home this budget is tight for every method. ForeWarn keeps 0.046 and the baselines 0.064 to 0.239.

W6 and Q5. Small test set, noise, paired confidence intervals, and breakdown by source.

The recall lead over StreamSense is outside noise on the falls and EDF, and the difference is within noise on OCCU and OOPS (Appendix H.6).

  • Paired bootstrap, macro over domains. The recall difference against StreamSense is +0.197, with an interval from +0.138 to +0.257. The clip-FAR difference is +0.071, and its interval includes zero.
  • Exact sign tests on discordant hazard clips. The falls give 10 against 1 with p = 0.012, and EDF gives 21 against 1. OCCU and OOPS give p = 0.45 and p = 0.75. On Home ForeWarn recalls more with p = 0.0001, at a higher clip FAR.
  • By source (Appendices H.6 and H.11). On the falls both methods catch all 10 CAUCAFall clips, and the lead comes from GMDCSA-24 with 19 of 21 against 10 of 21. On EDF ForeWarn is ahead for each of the five recorded subjects. Leaving out any one of the 64 OOPS source compilations keeps ForeWarn's recall between 0.873 and 0.892.
  • Unrelated patterns. Training and evaluation videos are separated by subject on the falls and by source channel on OOPS. With no training, the zero-shot scout reaches 0.774 and 0.827 on the falls and OOPS in the same controller. The immediate threshold was fixed from a sweep on EDF recordings that are also evaluated, and Appendix H.2 reports the sweep on all five domains. The held-out normal pools hold 16 to 88 clips, so the clip-FAR intervals are wide.

W7 and Q4. Lead time paired on events both systems detect, timely recall, and configurations that mitigate backlog.

On events both systems detect, ForeWarn is earlier than StreamSense on the falls and EDF and later on OOPS (Appendix H.5).

  • Paired lead. On the falls ForeWarn is earlier in 15 of 19 common hits, with a median of +0.50 s. On EDF it is earlier in 14 of 18, and on Home it is earlier in 12 and tied in 20 of 32. On OOPS it is later in 41 of 62, where StreamSense alerts on 0.784 of the normal clips.
  • Lead time. ForeWarn's median lead is 1.08 s on the falls, 0.86 s on EDF, 2.01 s on OOPS and 7.82 s on Home.
  • Timely recall. The recall R in the paper is timely recall. An alert counts only if its delivered time precedes the incident, and the delivered time includes the method's latency. Appendix H.12 and the supplementary material list the per-event times.
  • Backlog mitigation (Appendix H.9). In a new control experiment on the 64 GB tier, moving the verifier off the board to a separate server raises timely hits from 70 to 176 of 299, with 0.519 of normals alert-free against 0.803. Skipping stale steps gives 87 and a 1 s scout step 75 to 77.
AuthorsOpenReview ↗

Response to Reviewer bA3r (3/3)

W8. R65 is at 1 of 35, there is no GPU scheduling, and 95 per hour is disruptive.

The revision adds device experiments with three scheduling options and reports false warnings as distinct episodes (Sections 4.2 and 4.5, Table 9, Appendix H.9).

  • R65 at 1 of 35. On the two larger tiers these are decision misses and not scheduling losses. No advisory opened on any hazard clip, and the immediate path fired after the incident on nearly all alerted clips (Section 4.5).
  • GPU scheduling. We now evaluate three scheduling options in replay and on the device, as summarized under W7.
  • 95 per hour. Counted as distinct warning episodes the figure is 91 per hour on Home, against 38 for SPRINT, 82 for StreamSense and 30 for the guard. The revision uses these counts for the comparison (Table 9). Stricter budgets down to 12 per hour are in Appendix H.1 and summarized under W5.

Code. The supplementary material of the revision holds our release: the controller, the replay and calibration scripts, the data-role manifests, the cached scores, the per-clip decisions and the scripts that produce the tables of Appendix H. In the released manifests and decision files, clip identifiers that carry the folder names of recorded subjects are replaced by neutral codes, and local storage paths are removed. We have also located the fine-tuned scout weights, which the submission listed as missing. On the 64 GB board with the recorded input frames they reproduce the recorded scout scores step for step (Appendix F.4). The weights are not part of the release.

Reviewer_CEqgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the question of whether to issue an immediate alert or wait for further confirmation when a risk is predicted in streaming video. The authors propose the "ForeWarn" framework, which enables a model to decide—based on predicted future risks—when to trigger an alert and when to invoke a more powerful model for verification. A fine-tuned VLM is used to predict accident risks across different time horizons: high-risk events in the near term trigger immediate emergency alerts, while risks over a longer time horizon—combined with danger-status assessments from an asynchronous verifier—trigger advisory alerts. The effectiveness of this framework is validated across various real-world scenarios.

Critical strengths
  1. Assigning the tasks of risk prediction and risk confirmation to distinct decision-making roles is a meaningful approach to problem modeling. Confirming an existing risk is fundamentally different from predicting an impending accident; enforcing a strictly sequential confirmation process could negate the time advantage gained through prediction. A design that allows urgent risks to bypass verification—while maintaining a confirmation pathway for slower-developing events—is highly valuable.
  2. The evaluation incorporates datasets from diverse domains, ensuring a robust experimental setup. ForeWarn achieves SOTA macro-recall on public-domain datasets, with a computational cost comparable to other existing methods.
Critical weaknesses
  1. The effectiveness of the control rules lacks ablation validation. It remains unclear where the performance gains primarily originate—whether from the specifically fine-tuned "scout," the search thresholds, or the scheduling control rules as a whole.
  2. The performance improvements come at the cost of a significant trade-off involving false positives. In the OCCU and Home scenarios, ForeWarn’s balanced accuracy falls below the 0.500 baseline (which represents a "never-alert" strategy). This suggests that while the recall may be high, the system could impose a substantial burden in terms of excessive alerts during deployment. Furthermore, its real-time performance on actual hardware is suboptimal, falling short of what would be required for an effective early-warning system in edge-based streaming deployments.
Questions for authors

1.Would the results differ if other methods were applied using the specialized, fine-tuned "scout" described in the paper? It is recommended to include relevant ablation studies to address this. 2.In a new deployment, calibration using standard video data is required; while this helps reduce the burden of false alarms, does it compromise the system's early-warning capability in the new environment? How does the system handle cases where the "emergency path" itself exceeds the false-alarm budget? Furthermore, does the proposed approach offer any advantages over other baselines in these newly deployed scenarios?

5 comments
AuthorsOpenReview ↗

Response to Reviewer CEqg (1/2)

Thank you for the review. The ablations you asked for are now in Section 4.3 and Appendices H.3, H.4 and E.5. References below are to the revised paper, where changes are in blue. R / F denotes pre-incident recall and the clip false-alert rate (FAR) on held-out normal clips, and α is the nominal clip-FAR budget of site calibration.

W1. Where do the gains come from: the fine-tuned scout, the searched thresholds, or the control rules?

Each layer contributes in a different domain. The layers interact, so the revision reports the effect of changing one layer at a time and not shares (Appendix E.5).

  • Scout adaptation raises recall by 0.161 on the falls and by 0.262 on EDF with the same controller. On the falls its effect is largest at low false-alarm rates. At α = 0.1 a single threshold on the fine-tuned scout recalls 0.903 there, against 0.548 with the zero-shot scout (Appendix H.4).
  • Control structure. The immediate path alone gives 0.871 / 0.000 on the falls and 0.905 / 0.000 on EDF, so short-runway recall comes through that path. The advisory path adds 0.458 recall on Home and 0.120 on OOPS at a higher clip FAR. When every alert waits for the verifier, 7 of 31 clips are lost on the falls and 19 of 42 on EDF (Section 4.3).
  • Scheduling and calibration. On the long-runway home pool Home-L the risk-driven schedule recalls 0.630 with 7.22 verifier calls per clip, against 0.611 with 21.55 calls at a fixed 1 s cadence (Table 2). In a second replay the saving in calls holds and the fixed cadence recalls more (Appendix H.3). Site calibration raises Home-L recall by 0.204 at a specificity cost of 0.125.

W2a. Balanced accuracy below 0.5 on OCCU and Home.

OCCU and Home are the two hardest domains, and no method exceeds a balanced accuracy of 0.548 on either. Section 4.2 of the revision now describes both.

  • OCCU. The end of each fall is hidden behind an occluder. The four baselines recall at most 0.048 there, and ForeWarn recalls 0.119. The scout's short-horizon risk has an AUROC of 0.568 on OCCU against 0.858 on the staged falls (Appendix H.8).
  • Home. Home consists of clips from smart-home cameras with varied hazards, about half of which develop over a long runway. ForeWarn recalls 0.486 there, against at most 0.330 for the four baselines at their native operating points, at a higher clip FAR. Most of its false warnings on Home come through the advisory path (Appendix E.1).

In both domains ForeWarn's clip FAR exceeds its recall at the default budget, which puts its balanced accuracy below 0.5. On Home recall and FAR fall together with the budget (Appendix H.1).

W2b. Real-time performance on hardware.

In a new control experiment on the 64 GB tier, running the verifier off the board on a separate server raises timely hits from 70 to 176 of 299. The share of alert-free normals falls from 0.803 to 0.519 (Appendix H.9). Short-runway events also need a scout step that is faster than the available lead of about 1 s, which Section 4.5 states as a requirement. On the two larger tiers the controlled recordings stay at 1 to 2 of 35 under every schedule, because those misses are decision misses and not lag.

Q1. Would other methods improve with the same fine-tuned scout?

The fine-tuned scout is part of ForeWarn, and the one baseline that can take it already runs on it. StreamSense uses the same fine-tuned scout and the same 8B verifier and still trails on the falls and EDF.

  • StreamSense. Re-thresholded to ForeWarn's clip FAR it reaches 0.710 on the falls and 0.500 on EDF, against 0.935 and 0.929. On OOPS and Home the two methods are within three clips of each other (Table 10). Two of its rows in the submission had used the zero-shot scout. The revision corrects them, and its macro recall is now 0.473 (Table 1).
  • Scout alone. At α = 0.5 a single calibrated threshold on the fine-tuned scout, without the verifier, reaches ForeWarn's recall on the falls, EDF and OOPS, at a higher clip FAR on the falls and EDF. It recalls more on OCCU. At tight budgets the verified path adds recall on long-runway footage. On Home-L at α = 0.3 ForeWarn gives 0.370 / 0.432 against 0.204 / 0.318 (Appendix H.3).
  • SPRINT, PaSBench and the guard query the 8B with their own prompts and take no scout score, so the scout cannot be swapped into them.
AuthorsOpenReview ↗

Response to Reviewer CEqg (2/2)

Q2a. Does calibration on normal video compromise warning capability at a new site?

Calibration needs only normal footage, and where the precursor is distinctive it costs little recall. Tightening the budget from 0.5 to 0.1 lowers recall from 0.935 to 0.903 on the falls and leaves it at 0.929 on EDF (Appendix H.1). On Home recall comes through the advisory path, so it follows the budget and falls from 0.486 to 0.028 over the same range. Calibrating on the site is still better than carrying thresholds over. Relative to thresholds transferred from the development footage, site calibration raises OCCU recall by 0.095 at unchanged specificity and Home-L recall by 0.204 at a specificity cost of 0.125 (Section 4.3). A gate carried over from another domain gives a held-out clip FAR of 0.838 to 0.973 on OOPS (Appendix H.8). We therefore recommend one calibration pass on each site's own normal footage before use.

Q2b. What if the immediate path alone exceeds the false-alarm budget?

The revision adds a calibration-only rule that sets the immediate threshold to the lowest grid value that fits the budget on the calibration normals (Appendix H.2). This case occurs on OOPS at α ≤ 0.2. The new rule selects 0.93 at α = 0.2 and 0.97 at α = 0.1, which gives 0.640 / 0.216 and 0.373 / 0.054, respectively. The higher threshold trades short-runway recall for the budget.

Q2c. Is there an advantage over baselines in newly deployed scenarios?

On an additional set of 259 clips, with thresholds calibrated on its own normal clips, ForeWarn is ahead of a single threshold on the same scout at budgets of 0.2 and 0.3 (Appendix H.10). It reaches 0.309 / 0.282 and 0.414 / 0.436 there, against 0.171 / 0.205 and 0.343 / 0.410. The four baselines were not run on this set. Its industry and health groups are domains outside the scout's training sources, and ForeWarn recalls 0.538 and 0.200 of their hazards at α = 0.5. The controlled recordings are a further new site, and their result is given under W2b.

AuthorsOpenReview ↗

Follow-up to Reviewer CEqg

Dear Reviewer CEqg,

Thank you again for your review. Our response is posted above, together with a revised PDF. The ablations you asked for are now in Section 4.3 and Appendices H.3, H.4 and E.5. They separate the fine-tuned scout, the control rules and the calibration. The revision also compares StreamSense on the same fine-tuned scout at matched false-alert rates (Table 10) and reports calibration at a new site (Appendices H.1, H.2 and H.10).

Could you let us know whether these changes address your concerns, or which points remain open? We will answer any further question before the discussion closes on November 18. If the concerns are resolved, we would be grateful if you could reflect this in your assessment.

Best regards,

The authors

Reviewer_8SscOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ForeWarn, a risk-adaptive inference framework for proactive safety warning in streaming video. The key idea is to use predicted future risk at two different temporal horizons to determine both when to issue warnings and when to request more expensive verification. Specifically, the framework consists of an immediate warning path that directly acts on short-horizon risk predictions without waiting for verification, and an advisory path that combines longer-horizon risk predictions with verifier-confirmed hazardous states. A risk-dependent scheduling mechanism further controls the timing of repeated verification requests, while per-deployment calibration uses normal footage to determine advisory thresholds without requiring hazard labels.The paper also provides a one-step theoretical analysis of the value of verification and evaluates the proposed framework.

Critical strengths
  1. A meaningful problem formulation and intuitive risk-adaptive warning mechanism. The paper identifies an important limitation of conventional confirmation-based warning systems: waiting for stronger-model verification can cause early warning opportunities to be missed. The proposed separation between immediate forecast-based warnings and verifier-confirmed advisories is intuitive and well motivated. Furthermore, using predicted future risk to guide both warning decisions and verification scheduling provides a coherent framework for balancing early detection, confirmation quality, and inference cost.

  2. Comprehensive experimental evaluation with practical deployment considerations. The paper evaluates ForeWarn across five public video domains and includes additional live trials and experiments on three Jetson edge-device configurations. The ablation studies investigate the contributions of the two warning paths, verification scheduling, and per-deployment calibration. Importantly, the authors also account for inference latency and accumulated processing delays rather than relying exclusively on offline replay results. This provides a relatively comprehensive assessment of the proposed system and exposes important practical limitations of streaming safety warning.

Critical weaknesses
  1. Limited theoretical and calibration guarantees. The one-step analysis provides useful intuition but does not fully characterize the sequential, asynchronous decision-making process involving repeated verification, persistent states, and evolving risk estimates. Similarly, the main per-deployment calibration relies on a joint threshold search without formal false-alert guarantees. Although the paper includes a fixed-score conformal alternative and empirical calibration analysis, the reliability of the complete policy under distribution shifts and extended deployments remains insufficiently established.

  2. High false-alarm burden and limited practical deployability. The immediate warning path bypasses verification, improving early detection at the cost of potentially excessive false alarms. ForeWarn reports approximately 95 false alerts per hour on held-out home normal videos, raising serious concerns about its suitability for continuous monitoring. Moreover, the edge-device experiments show that accumulated processing delays substantially reduce timely recall compared with replay evaluation. These findings suggest a considerable gap between the reported benchmark improvements and the requirements of reliable real-world safety warning systems.

Questions for authors

Q1. Sequential decision analysis. Can the authors clarify how the one-step verification criterion relates to the complete sequential policy? In particular, how do repeated queries, delayed responses, and persistent WATCH states affect the validity of the theoretical conclusions?

Q2. False-alarm burden. Given the approximately 95 false alerts per hour on home normal videos, what operating points are achievable under substantially stricter false-alarm budgets? Can the authors provide recall–false-alert trade-offs that better reflect continuous real-world monitoring requirements?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8Ssc

Thank you for the careful reading. References below are to the revised paper, where changes are in blue. R / F denotes pre-incident recall and the clip false-alert rate (FAR) on held-out normal clips, and α is the nominal clip-FAR budget of site calibration.

W1a and Q1. One-step analysis and the sequential, asynchronous policy.

Proposition 1 covers a single verification decision. The revision states this scope in Section 3.5 and reports how the sequential policy behaves on the three aspects you name (Appendix H.7).

  • Repeated queries. On OOPS, 30 of the 43 advisory hits are opened by the first query. On Home, 25 of 53 are opened by the third or a later query. These later queries are outside the one-step analysis (Appendix F.2).
  • Delayed responses. The one-step quantity includes the event that the answer returns before the incident. In replay few answers return after it, for example 12 of 88 on the falls. In the recorded device runs this share reaches 26 of 26 on the falls, because the median verifier latency there is 2.0 to 2.9 s against the 0.483 s charged per call in replay. This is the regime in which the proposition says verification cannot help.
  • Persistent WATCH states. An advisory is emitted only on entry to WATCH, so persistence does not repeat a warning but re-entry does. WATCH episodes last a median of 3 to 18 steps and occur 0.16 to 0.91 times per clip.

W1b. Guarantee of the joint threshold search, and reliability under shift and extended deployment.

The guarantee of Proposition 2 holds for a scoring rule fixed before calibration. The joint search is assessed empirically, and the revision reports its calibration and held-out rates side by side (Appendix H.8).

  • At the nominal budget of 0.5 the held-out clip FAR is 0.125, 0.143 and 0.286 on the falls, EDF and OCCU. On OOPS and Home it is 0.676 and 0.580, so the revision calls α a nominal budget (Section 3.4).
  • Under shift, a gate carried over from another domain gives a held-out FAR of 0.838 to 0.973 on OOPS and 0.466 to 0.636 on Home. This is why the method calibrates on each site's own normal footage.
  • The longest held-out normal pool is 0.82 h, so extended deployment is outside the evaluation (Appendix F.3).

W2a and Q2. About 95 false alerts per hour, and operating points under stricter budgets.

Under strict budgets ForeWarn keeps most of its recall on the staged falls and EDF, where the precursor is distinctive. On Home recall follows the budget, and the revision reports the full range of operating points (Appendix H.1).

  • Falls and EDF. At α = 0.1 ForeWarn's held-out points are 0.903 / 0.062 on the staged falls and 0.929 / 0.143 on EDF. Chosen on calibration normals for at most 12 false warnings per hour, it gives 0.690 on EDF with no false warning and 0.839 on the falls with one false warning step.
  • Home. Recall on Home comes through the advisory path, so it falls with the budget. It goes from 0.220 / 0.341 at α = 0.3 to 0.028 / 0.080 at α = 0.1, with 54 and 10 false-warning steps per hour.
  • OOPS. With the immediate threshold set from calibration normals, ForeWarn gives 0.640 / 0.216 at α = 0.2 and 0.373 / 0.054 at α = 0.1 (Appendix H.2).
  • Per-hour count. The 95 per hour of the submission counted warning steps on Home. Counted as distinct warning episodes the figure is 91 per hour, against 38 for SPRINT, 82 for StreamSense and 30 for the guard. The revision uses these counts for the comparison (Section 4.2, Table 9).

W2b. Accumulated delay on edge devices, and the gap between benchmark and real-world requirements.

On the 64 GB tier, moving the verifier off the board recovers the timely hits on the public set, at a cost in specificity (Appendix H.9).

  • Public set, 64 GB tier. With the verifier on a separate server, timely hits rise from 70 to 176 of 299, and the share of alert-free normals falls from 0.803 to 0.519. Skipping stale steps and a 1 s scout step help less, with 87 and 75 to 77 timely hits.
  • Controlled recordings. On the two larger tiers the controlled recordings stay at 1 to 2 of 35 under every schedule. Their lag is small and they stay there with the verifier off the board, so these are decision misses (Section 4.5).
  • Requirement. For short-runway events the scout's latency per step must stay below the lead that the immediate path provides, about 1 s on the falls. Section 4.5 states this requirement.
AuthorsOpenReview ↗

Follow-up to Reviewer 8Ssc

Dear Reviewer 8Ssc,

Thank you again for your review. Our response is posted above, together with a revised PDF. The revision states the scope of Proposition 1 and reports the sequential behavior of the policy (Section 3.5, Appendix H.7). It also adds operating points under stricter false-alert budgets (Appendix H.1) and a device experiment on verification scheduling (Appendix H.9).

Could you let us know whether these changes address your concerns, or which points remain open? We will answer any further question before the discussion closes on November 18. If the concerns are resolved, we would be grateful if you could reflect this in your assessment.

Best regards,

The authors

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion and the revision

Dear Area Chair and Reviewers,

Thank you for the reviews and the discussion. This comment summarizes the revised paper and lists where each review is answered. Changes in the revised PDF are in blue, and the code is in the supplementary material.

Status of the discussion. Reviewer bA3r read the revision and raised the rating from 2 to 4. Reviewers h1sV and CEqg replied and keep their ratings. Reviewer h1sV raised no further question, and our answer to the three remaining points of Reviewer CEqg is posted in that thread. We have not yet heard from Reviewers 8Ssc and Mnop, and we thank the Area Chair for inviting the reviewers to respond. The list at the end maps each review to the revision.

Main results after the revision.

  • Short-runway hazards. ForeWarn warns from the scout on the short horizon and does not wait for the verifier. In a re-scored replay with the same scout and verifier, a controller that makes every alert wait for the verifier loses 7 of 31 staged falls and 19 of 42 EDF hazards (Section 4.3, Appendix H.4). In the main evaluation, StreamSense on the same scout and verifier reaches 0.710 and 0.500 on these two domains against 0.935 and 0.929 once it is re-thresholded to ForeWarn's clip FAR (Table 10). A budget of 12 false warnings per hour allows none on these two domains. In that post-hoc comparison the immediate path alone keeps 0.871 and 0.905 of the hazards, against 0.387 and 0.381 for StreamSense and at most 0.194 for the other baselines (Appendix H.1).
  • Statistical support. At the operating points of the main results, the macro recall difference against StreamSense is +0.197, with a paired bootstrap interval from +0.138 to +0.257. The clip-FAR difference is +0.071, and its interval includes zero (Appendix H.6). At matched clip FAR the lead is 13.3 points. It comes from the staged falls and EDF, and on OOPS and Home the two methods are close (Section 4.2, Table 10).
  • Long-runway hazards. At tight budgets the verified advisory path adds recall. On the long-runway home pool Home-L at α = 0.3, ForeWarn reaches a recall of 0.370 at a clip FAR of 0.432. A single threshold on the same scout reaches 0.204 at 0.318 (Appendix H.3).
  • Edge devices. On the 64 GB tier, moving the verifier off the board to a separate server raises timely hits from 70 to 176 of 299. The share of alert-free normal clips falls from 0.803 to 0.519, and our controlled recordings stay at 1 or 2 timely hits of 35 under every schedule (Appendix H.9).
  • Scope. The advantage of ForeWarn is on short-runway hazards with a distinctive precursor. On Home, recall follows the alert budget, from 0.028 at α = 0.1 to 0.486 at α = 0.5 (Table 15). Under the budget of 12 per hour ForeWarn keeps 0.046 there against 0.064 to 0.239 for the baselines (Appendix H.1). On OCCU the end of each fall is hidden behind an occluder. The four baselines recall at most 0.048 there, and ForeWarn recalls 0.119 at a clip FAR of 0.286. On both domains ForeWarn's clip FAR exceeds its recall at the default budget, and no method exceeds a balanced accuracy of 0.548 on either (Section 4.2).

Where each review is answered.

  • Reviewer 8Ssc. Scope of the one-step analysis and behavior of the sequential policy (Section 3.5, Appendix H.7). Operating points under stricter false-alert budgets (Appendix H.1). Device experiment on verification scheduling (Appendix H.9).
  • Reviewer CEqg. Ablations that separate the fine-tuned scout, the control rules and the calibration (Section 4.3, Appendices E.5, H.3 and H.4). StreamSense on the same fine-tuned scout at matched false-alert rates (Table 10). Site calibration under tighter budgets, a calibration-only rule for the immediate threshold and an additional set of 259 clips (Appendices H.1, H.2 and H.10).
  • Reviewer bA3r. Single small VLM, single large VLM and vision-only arms (Section 4.3). Matched and strict false-alert budgets (Table 10, Appendix H.1). Paired lead times and paired tests (Appendices H.5 and H.6). Sample frames of the test videos (Figure 5).
  • Reviewer Mnop. What makes OCCU hard and the mitigations that we tested (Section 4.2, Appendices E.5, H.2 and H.4). Three scheduling options on the device (Appendix H.9).
  • Reviewer h1sV. Choice of the immediate threshold and a sweep over it on all five domains (Section 3.3, Appendix H.2). Direct comparison with cascades and confidence routing (Sections 2 and 4.3). Intervals for the main results (Appendix H.6). The Ethics Statement now describes who was recorded, consent and review status.

We thank the reviewers and the Area Chair for their time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.