{"id":"7bca4eb9-7ab1-42b3-8b18-6dff6baefdd8","arxiv_id":"2607.13569","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A grounded edge-cloud pipeline applies VLM reasoning only to localized passenger and farebox evidence, achieving 31–54% five-class zero-shot payment accuracy on two real bus videos.","lead":"A bus-video analytics system detects stops and passengers with lightweight edge models, then calls a vision-language model only on short passenger clips and farebox crops to classify payment method with no payment-specific training. On two real bus videos the five-way payment accuracy is 31–54%, so the system shows potential rather than production readiness.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim that grounding improves VLM reasoning is untested: no baseline against full-video or uncropped VLM prompting; current point estimates cannot support 'realizable'.","rationale":"The reader's conditional verdict is appropriate; I do not propose moving it. My concern is not internal inconsistency but an evidence gap that is more load-bearing than the reader's identified projective-prior assumption. The paper's title says 'realizable' and the abstract says grounding reduces cloud inference and supplies localized evidence. The experiments demonstrate that a specific instantiation yields non-trivial stop/passenger F1 and weak payment accuracies, but no comparison isolates the contribution of grounding. Stage 1/Stage 2 refinement tests only the coarse-to-fine sampling within grounded crops (Table 2: 0.349→0.313 on C3_1, 0.485→0.536 on C3_3), not whether grounding per se helps. The alternative explanation—that any short, sampled clip with a farebox crop would yield similar or better results—is not excluded. The reader's Eq. 5/9 concern is valid but would only affect clip selection precision; even with perfect selection, the central causal claim could remain unproven. The duration accounting inconsistency (486 vs 436 vs ~540 minutes) is a reporting flaw that should be fixed but is not the central argument. Therefore the verdict stays CONDITIONAL, pending an ablation/baseline check.","tokens_in":10217,"tokens_out":4508,"duration_ms":47502,"concrete_test":"Release the exact prompts and run the same GPT-5.4-mini payment classifier on four inputs per passenger on C3_3: (a) GHR-VLM Stage 2 contact sheet (current); (b) the same 16 frames from the passenger clip, full-frame, no farebox crop; (c) 16 uniformly sampled frames from the entire stop interval (no passenger clip); (d) the full passenger clip at native resolution without temporal downsampling (or a 16-frame sample if API limits). Compare five-class accuracy and per-passenger token count. If (b) or (c) matches or exceeds the 0.536 Stage 2 accuracy on C3_3, the localized-evidence claim fails; if (a) uses substantially fewer tokens than (d) with equal accuracy, then the cost claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim—that invoking the VLM only on grounded passenger clips and contact sheets 'reduces cloud inference' and 'supplies the localized evidence that VLMs otherwise struggle to identify'—is not actually tested. Section 4.3 compares Stage 1 vs Stage 2, both of which crop the farebox and use downsampled grounded contact sheets; there is no baseline in which the same VLM sees the full passenger clip, the full stop interval, or the raw video without grounding. Consequently the central causal assertion (explicit grounding improves VLM reasoning) is indistinguishable from the trivial fact that shorter inputs are cheaper. Token/cost savings are asserted but never quantified. The paper's own numbers are point estimates on two private videos with a duration inconsistency (486 min abstract vs 436 min §4.1 vs ~540 min '9 hours'), and no code or prompts are released, but the missing baseline is more load-bearing because even perfect stop/passenger grounding would not establish the headline contribution. The Eq. 5/9 projective-prior risk is real but secondary: it concerns precision of the pipeline, whereas the absent baseline concerns whether the architecture has any causal benefit at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"GHR-VLM proposes an edge-cloud framework for zero-shot payment classification on bus surveillance video. The edge pipeline detects stop intervals from door state, tracks passengers with SAM-based instance segmentation, and generates passenger clips; a VLM direction filter selects boarding passengers via a fixed orientation-to-behavior mapping; a two-stage farebox-cropped contact-sheet classifier assigns QR, cash, tap, swipe, or evade labels. Evaluation on two real videos reports stop F1 of 0.887/0.898, passenger-clip F1 of 0.702/0.847, five-class payment accuracy of 0.313/0.536, and binary evasion accuracy of 0.783/0.837. The paper claims that explicit visual grounding reduces cloud inference, avoids payment-specific training data, and provides the localized evidence that VLMs otherwise struggle to identify.","tokens_in":10443,"tokens_out":5487,"duration_ms":60190,"significance":"If the central claim were established, the framework would be practically valuable for transit agencies: it targets a real operational need, uses only real surveillance data, and proposes a plausible division of labor between lightweight edge modules and a cloud VLM. The paper is also honest about where the second stage hurts performance (C3_1) and about degraded video conditions. However, the core causal claim — that explicit grounding improves VLM reasoning — is not tested, because no comparison is made against an ungrounded VLM baseline or against a cost-equivalent full-input baseline. The empirical basis is also thin: two private videos, no confidence intervals, and no end-to-end passenger-level accuracy. The framework is promising, but the evidence does not yet support the 'realizable' language in the title and abstract.","major_comments":[{"comment":"The abstract and Sec. 1 claim that explicit visual grounding improves VLM reasoning and reduces cloud inference, but no experiment supports this. The only VLM comparison is Stage 1 vs Stage 2, both of which operate on farebox-cropped contact sheets built from the same grounded pipeline. There is no baseline in which the same VLM sees the full passenger clip, the full stop interval, or the raw video without grounding. Consequently, the observed accuracy differences are confounded with input-length reduction and cannot establish that the grounded spatiotemporal evidence itself helps. The paper should include such a baseline and also quantify token/cost savings, since 'reduces cloud inference' is an asserted benefit that is never measured.","section":"Sec. 4.3, Tables 1-2"},{"comment":"The pipeline assumes that the largest tracked box is the boarding passenger nearest the farebox (Eq. 5) and that a VLM orientation judgment maps deterministically to boarding/alighting/staying (Eq. 9: front/side → boarding, back → alighting, inside → staying). These assumptions are fragile in exactly the situations the paper itself acknowledges: the driver leaving the cockpit, simultaneous boarding events, and passengers reappearing (Sec. 4.2). When Eq. (5) selects the wrong identity or Eq. (9) filters out a true boarder, no passenger clip reaches the payment classifier, and the error propagates to the final output. The paper reports component-level stop and clip F1, but not the joint passenger-level payment accuracy. The authors should report the end-to-end success rate and analyze these failure modes quantitatively.","section":"Sec. 3.2-3.3, Eqs. (5) and (9)"},{"comment":"The empirical evidence is limited to two private videos with point estimates and no confidence intervals. The abstract states 486 minutes, while Sec. 4.1 says 436 minutes, a discrepancy that should be resolved. On C3_1, Stage 2 reduces five-class accuracy from 0.349 to 0.313 and binary evasion accuracy from 0.813 to 0.783; the paper attributes this to visual quality, but no statistical analysis is provided. Given the title's promise of making zero-shot transit video analytics 'realizable', the authors should either provide more data, per-stop or per-error breakdowns, or at least temper the claim to match the two-video, unstable results.","section":"Sec. 4.1, 4.3, Tables 1-2"}],"minor_comments":[{"comment":"The duration inconsistency (486 min vs 436 min vs '9 hours') should be corrected in the final version.","section":"Sec. 4.1"},{"comment":"Please define 'alternate frames' explicitly (e.g., every 2nd frame at 10 FPS) and clarify the relationship between sampled-frame indices and absolute timestamps in Eq. (6)-(7).","section":"Sec. 3.2"},{"comment":"The margin notation '(0.35,0.35,-0.50)' is unclear: are these absolute normalized offsets or fractions of the box dimensions? Please define the coordinate convention once and reuse it.","section":"Sec. 3.4"},{"comment":"When Stage 1 returns a non-contiguous evidence-frame set, Stage 2 uses the earliest and latest frames, which may include unrelated frames. Please describe how evidence frames are combined or justify this choice.","section":"Sec. 3.4"},{"comment":"No prompt templates, model version details, or code are provided. Given the emphasis on zero-shot behavior, releasing the exact prompts would be important for reproducibility.","section":"Sec. 4"},{"comment":"Class-wise counts are small and arrows indicate only relative changes; please provide per-class sample sizes or confidence intervals to help judge the stability of the reported improvements.","section":"Sec. 4.3, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a systems-style description with a plausible architecture, and the authors are candid about negative results. The main barrier is not correctness of the implementation but the absence of a decisive experiment that isolates the effect of explicit grounding. I would ask for the ungrounded-baseline experiment and cost measurement before considering acceptance. The duration inconsistency and the lack of end-to-end accuracy are also easy fixes. I do not see a circularity problem: the self-citations [10][11] are related prior work, not premises of the current derivation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kaicong — quick take on GHR-VLM.\n\nThe paper is a system for passenger-level payment classification from bus surveillance. It combines SAM3, a rule-based tracker, and GPT-4o/GPT-5.4-mini in an edge-cloud pipeline: ground the door, detect stops, track passengers, filter by orientation, then classify payment via cropped farebox contact sheets. The stop and passenger-clip grounding numbers are decent (F1 0.70–0.90) and the paper is honest about the weak payment-classification results (five-class accuracy 0.31–0.54, with Stage 2 hurting C3_1). That honesty is a real plus.\n\nWhat's genuinely new is the assembly: the two-stage coarse-to-fine contact-sheet refinement, asking the VLM to select its own evidence frames, and the CS mapping that turns a generic orientation judgment into boarding/alighting. None of the components are new, but the combination is a legitimate system for a real problem.\n\nThe soft spots, in order of size. First, the central claim is untested. The abstract says grounding 'supplies the localized evidence that VLMs otherwise struggle to identify,' but the only comparison is Stage 1 vs Stage 2, both of which crop the farebox and use downsampled contact sheets. There is no baseline where the same VLM sees the full passenger clip or the raw video without grounding. So the grounding-benefit claim is indistinguishable from the trivial fact that shorter inputs are cheaper. Second, no code, prompts, or data are released, and all numbers are point estimates on two private videos. Third, the duration accounting is inconsistent: 486 minutes in the abstract, 436 in §4.1, and '9 hours' in the same section. Fourth, the projective prior (largest box equals boarding passenger; orientation maps deterministically to direction) is fragile when people linger or board in groups — the paper itself mentions those cases. That's a real precision risk, but it's secondary to the missing baseline.\n\nThis paper is for researchers working on transit video analytics or on using VLMs for fine-grained activity classification under a grounding prior. It is a proof of concept, not a validation. It does not deserve a desk reject — the system is plausible and the honest reporting gives referees something to work with — but a serious referee should demand a naive-VLM baseline, released code/prompts, and a toned-down title before it's publishable. If the baseline is added and shows a real gain, it becomes a useful paper. If it shows no gain, that's also a useful negative result.\n\nI'd accept it for peer review, with the expectation of heavy revision.","headline":"A readable proof-of-concept system with honest reporting of weak payment results, but its central claim that grounding improves VLM reasoning is untested — no baseline against ungrounded prompting.","tokens_in":11000,"tokens_out":2710,"would_cite":false,"duration_ms":27668,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A VLM with no payment training can classify bus payment types when the video is pre-ground to passenger clips and farebox close-ups.","keywords":["transit video analytics","vision-language models","visual grounding","edge-cloud collaboration","zero-shot recognition","passenger payment classification","bus surveillance video"],"falsifier":"Run the pipeline on a stop where two passengers board together and where an alighting passenger remains beside the payment box; if passenger-clip F1 and payment accuracy do not fall compared with isolated single-boarder stops, then the largest-box and direction-assumption priors are not load-bearing. A simpler version: place a stationary dummy near the payment box while a passenger boards and check whether the clip sent to the classifier contains the wrong person.","tokens_in":10043,"feed_emoji":"🚌","tokens_out":11551,"duration_ms":103758,"temperature":0.7,"pith_summary":"The paper claims that a general-purpose vision-language model, with no payment-specific training, can classify how a bus passenger pays—QR, cash, tap, swipe, or evasion—if it is shown compact, passenger-centered evidence instead of raw surveillance footage. The design splits the work across edge and cloud: lightweight edge modules detect door openings to isolate stops, track passengers, and crop passenger clips, while a cloud VLM first judges each passenger's facing direction and then classifies payment on a farebox-cropped contact sheet, refining with a second pass built from the frames the VLM itself selected as evidence. On real-world bus video, the pipeline reaches stop-detection F1 around 0.89, passenger-clip F1 around 0.70–0.85, and five-class payment accuracy around 0.31–0.54, with the worse numbers on the less consistently lit video. The paper's central point is that explicit spatiotemporal grounding is what makes zero-shot transit video analytics practical, and that the same grounding exposes how strongly fine-grained recognition still depends on video quality.","feed_headline":"Grounding clips to the farebox yields zero-shot bus-payment labels","feed_subtitle":"A lightweight edge tracker cuts each stop to a short clip, letting a general VLM sort QR, cash, tap, swipe, and evasion.","key_machinery":"The load-bearing mechanism is spatiotemporal grounding, built from four pieces: (1) door-grounded stop filtering, which uses the front door's open/close state to cut the video into stop intervals; (2) rule-based passenger tracking with a projective prior—the largest tracked box is assumed to be the passenger nearest the farebox; (3) a complex-to-simple (CS) direction mapping that reduces boarding/alighting/staying to a generic facing-direction judgment (front/side→boarding, back→alighting, inside→staying); and (4) a two-stage coarse-to-fine payment classifier that crops a farebox region into a 16-frame contact sheet, asks the VLM to select its own evidence frames, then refines on a tighter c","core_discovery":"GHR-VLM claims that invoking a VLM only on grounded passenger clips and farebox contact sheets makes zero-shot passenger-level payment classification realizable on real bus surveillance video. A door-grounded edge pipeline first filters the long stream down to stop intervals, then a rule-based tracker follows passengers and selects the largest tracked box as the boarding passenger nearest the farebox. A 'complex-to-simple' mapping lets the VLM judge only whether the person faces front, side, back, or inside, and a fixed rule converts those judgments to boarding, alighting, or staying. Retained clips are then submitted to two VLM passes over farebox-cropped contact sheets: the first samples s","pith_inferences":["The same grounding recipe—a fixed geometric prior, edge-side clip extraction, and an evidence-crop refinement—should transfer to other 'needle-in-a-haystack' fine-grained video tasks such as retail checkout verification or package handling, wherever a stable camera supplies a localization prior.","Because the pipeline assumes the largest tracked box is the boarding passenger, scenarios with group boardings, alighting passengers lingering by the farebox, or a driver leaving the cockpit will systematically corrupt the clip sent to the classifier; a stress test varying these scene factors would map the failure envelope.","At five-class accuracy of 0.31–0.54, full payment-method auditing is not yet production-ready; the binary evade/non-evade outputs (accuracy ~0.78–0.84) are closer to operationally usable and may be the right first deployment target.","Stage 2 refinement helped the well-lit video and hurt the poorly lit one, which suggests a cheap adaptive rule—keep the coarse prediction when visual quality or VLM confidence is low—could improve overall robustness."],"forward_implications":["Zero-shot transit payment analytics no longer requires payment-specific annotations, so the pipeline can be deployed to a new bus fleet, camera view, or agency without retraining.","Cloud inference is reduced to a few short VLM calls per stop, because only passenger clips and contact sheets leave the edge.","The two-stage evidence refinement can improve fine-grained recognition on well-lit footage—Swipe recall on the nighttime video rises from near zero to 0.333—showing that focused evidence helps a general VLM recover subtle actions.","The pipeline outputs structured stop, passenger, direction, and payment events, which transit agencies can use for crowding management, service planning, and revenue auditing.","The CS mapping means a general-purpose VLM does not need transit-specific action knowledge; it only needs to judge facing direction, a generic visual skill."],"fun_headline_variants":["Door-grounded clips let VLMs label bus payments zero-shot","Edge-cloud GHR-VLM turns bus video into farebox proof","Zero-shot payment labels via grounded passenger clips","Hybrid reasoning grounds VLMs for transit video analytics","Bus pay analytics without training data: GHR-VLM"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole chain rests on assuming the largest tracked box in the door area is the passenger who just boarded and is about to pay; if someone alighting lingers by the payment box, passengers board in a group, or the driver steps out, the payment classifier receives a wrong clip or no clip at all.","fun_headline_variants_meta":{"raw":{"variants":["Door-grounded clips let VLMs label bus payments zero-shot","Edge-cloud GHR-VLM turns bus video into farebox proof","Zero-shot payment labels via grounded passenger clips","Hybrid reasoning grounds VLMs for transit video analytics","Bus pay analytics without training data: GHR-VLM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000146,"raw_usage":{"total_tokens":1020,"prompt_tokens":748,"completion_tokens":272,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":192}},"tokens_in":492,"tokens_out":272,"duration_ms":3749,"temperature":1.0,"reasoning_tokens":192,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T04:47:01.782561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a stop where two passengers board together and where an alighting passenger remains beside the payment box; if passenger-clip F1 and payment accuracy do not fall compared with isolated single-boarder stops, then the largest-box and direction-assumption priors are not load-bearing. A simpler version: place a stationary dummy near the payment box while a passenger boards and check whether the clip sent to the classifier contains the wrong person.","supporting_citations":[],"review_version":1}