{"id":"0a6ece80-d3d5-4c2d-837d-c138098d1a2a","arxiv_id":"2508.20687","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"diveXplore 6.0 adds one-second shot sampling, alternate map search, OCR, speech-to-text, and temporal search for VBS2022, without a reported evaluation.","lead":"This paper describes diveXplore 6.0, an interactive video search system prepared for the Video Browser Showdown 2022. It lists planned upgrades, such as one-second shot sampling and speech-to-text, but provides no measurements showing the upgrades work.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central performance claims (e.g., 1s segmentation improves KIS) rest on unmeasured assumptions; no evaluation supports Table 2, so the system description is plausible but unverified.","rationale":"The paper is a VBS system description. Its value is in describing system changes, not in proving them. The central claim as identified by the reader is that 1s segmentation improves KIS retrieval. This is an empirical claim, and the paper contains no experiment, no ablation, no comparison with version 5.0, and no VBS 2022 results. The absence of evidence is not a formal inconsistency, but it means the claim is unverified. The reader's verdict of UNVERDICTED is therefore appropriate: the description is plausible and internally coherent, but the load-bearing assumptions about cause and effect are not demonstrated. I do not find a deeper internal flaw in the architecture description; the main gap is missing validation. Thus no change to the reader's verdict is needed.","tokens_in":4332,"tokens_out":1870,"duration_ms":18674,"concrete_test":"Run a controlled KIS experiment on the V3C1/V3C2 test set: keep the diveXplore 5.0 feature pipeline and interface, vary only the shot segmentation (10s vs 1s), and measure KIS success rate and time over the same task set (e.g., 20 known-item tasks with short target segments). If success rate does not improve (or time to find increases), the §2.3 attribution fails. Also independently re-run with ranking/feature changes disabled to isolate the segmentation effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 and Table 2 assert that reducing the keyframe sampling interval from 10s to 1s 'improves KIS performance,' attributing missed KIS items to the 10s shot length. The load-bearing assumption is that the segmentation interval, rather than feature vocabulary, ranking, or interface, causes the misses. No KIS evaluation is reported; the paper's own language is hedged ('seems to greatly impede'). If short target segments are missed because their keyframes are not discriminative under the Inception/YOLO/Places features, or because ranking buries them among 10x more shots, 1s sampling would not yield the claimed improvement and could degrade precision by multiplying candidate shots. Similarly, the 'temporal search' improvement in Table 2 is never actually specified in §2.3, so there is no concrete mechanism to test. Without ablation or VBS-style task results, the central claims are unsupported assertions rather than findings.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes diveXplore 6.0, ITEC's interactive video retrieval system for the VBS 2022 competition. It presents the system's modular web-based architecture, two search modes (shot and map search), the set of deep features (Inception v3 concepts, YOLO objects, Moments in Time events, Places), and the user interface. The main contribution is Section 2.3 and Table 2, which list four improvements over diveXplore 5.0: reducing the keyframe sampling interval to one second, an alternative map-search mechanism, adding OCR/STT, and temporal context search. The paper contains no task-level evaluation or quantitative comparison with the previous version; the improvement claims are expressed in hedged, qualitative terms.","tokens_in":4539,"tokens_out":6125,"duration_ms":58316,"significance":"If the claims were substantiated, the paper would provide a useful data point for interactive video retrieval: the move to 1s shot sampling and the integration of temporal context features could inform other VBS systems. The paper also offers a clear description of a deployed, working interactive search system, including UI details and feature choices. However, none of the improvements are validated: there are no ablation results, no offline retrieval experiments, and no VBS 2022 outcome data. The paper is therefore best regarded as a system description with design hypotheses, not as an evaluation-driven contribution. The authors are transparent about the VBS 2021 rank (12/18), which is a strength, but the significance of the proposed changes remains unknown.","major_comments":[{"comment":"The central claims in Table 2—especially '1s segmentation ... Improving KIS performance'—are not supported by any evaluation. The only evidence cited is the hedged observation in §2.3 that 'the rather long 10s shot segmentation seems to greatly impede KIS tasks.' No VBS task-level results, no offline KIS experiment, and no comparison of diveXplore 5.0 vs. 6.0 are reported. As written, the improvement is a design hypothesis, not a finding.","section":"§2.3, Table 2"},{"comment":"Even if one accepts that KIS misses occur, the paper does not establish the sampling interval as the cause. Missed targets could be due to non-discriminative features (Inception/YOLO/Places), to ranking that buries short segments, or to the map/shot interface. Moreover, reducing the interval from 10s to 1s multiplies the number of shots by about ten, which can hurt precision and increase browsing time. The paper should either provide an ablation isolating the sampling interval or qualify the claim as a hypothesis to be tested at VBS 2022.","section":"§2.3, Table 2"},{"comment":"Two of the four listed improvements—'alternative map search' and 'temporal search'—are not actually described. The text says only 'we explore alternative ways to match those features' and 'the system requires more sophisticated means for temporal context search, which also will include single keyframe-based features.' No retrieval mechanism is given, and no precision/recall measurement supports 'Improving result precision.' Please specify the algorithms and provide at least a baseline comparison, or remove these entries from the table.","section":"§2.3, Table 2"},{"comment":"The Abstract states that the v5.0 refactoring 'proved to be a sensible decision' because 'the new system showed increasing performance in VBS2021,' but §1 reports only the absolute rank 12/18. Without prior ranks (VBS2017–VBS2021) and ideally task-level scores, 'increasing performance' is unsubstantiated; VBS rank also depends on the number and skills of competing teams. Provide the underlying numbers or soften the claim.","section":"Abstract and §1"}],"minor_comments":[{"comment":"Please clarify whether '1s segmentation' changes shot boundaries or only the keyframe sampling interval. Section 2.1 says shots are 'ten-second scenes' and that events are computed on the full ten-second segments; Section 2.3 uses 'shot segmentation' and 'keyframe sampling interval' interchangeably.","section":"§2.1, §2.3"},{"comment":"The phrase 'which also will include single keyframe-based features' is ambiguous: does temporal search combine segment-level event features with per-second keyframe features? Please define the temporal context search pipeline.","section":"§2.3"},{"comment":"Minor grammatical issue: 'As depicted the front end part of Figure 1' should be 'As depicted in the front-end part of Figure 1.'","section":"§2.2"},{"comment":"The 'Impact' column values such as 'Improving KIS performance' are outcome claims; consider renaming the column 'Intended effect' to distinguish design goals from measured outcomes.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"This is a VBS system-description paper. For a research journal, the missing evaluation is significant; the central claims in Table 2 need support or rephrasing. If the venue accepts short system demonstrations without experiments, a lighter revision may suffice; otherwise the authors should add a post-VBS 2022 results section or an offline ablation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You can read this in ten minutes. It's a five-page competition system description from the diveXplore group, reporting what they changed for VBS 2022: one-second shot sampling instead of ten, alternative map matching, OCR and STT, and something they call temporal search. None of these claims are backed by any measurement. That is the whole story.\n\nWhat's actually new: the specific 6.0 configuration is new, and the paper describes the architecture, interface, and feature set clearly enough for someone in the VBS orbit to understand what was built. The observation that the ten-second segmentation 'seems to greatly impede' KIS tasks is plausible, and reducing the sampling interval to one second is a reasonable engineering response. Credit is due for honesty: the paper uses hedged language and doesn't present unverified guesses as findings.\n\nThe soft spot is not subtle. Table 2 lists '1s segmentation' and 'temporal search' as improvements, but no evaluation is reported. The single rank from VBS2021 (12/18) has no context. The load-bearing assumption in Section 2.3 is that the sampling interval, not the features or ranking, caused the missed KIS items; the paper gives no evidence for that. Also, 'temporal search' in Table 2 is never actually described in Section 2.3, so there is no mechanism to evaluate. A one-second interval will multiply the number of candidate shots by an order of magnitude, which could hurt precision and response time just as easily as it helps recall; nobody can tell from this text.\n\nThis is a system description for a competition, not a research result. For someone building a VBS system, it's a useful data point about what one experienced team tried. For anyone expecting a finding, there is nothing here.\n\nI would send it to a short-paper or workshop track if the venue expects system descriptions, and ask the authors to include even a minimal report of VBS2022 outcomes or a small ablation. As a journal contribution it would need real evaluation. Worth a serious referee for the workshop; not because it proves anything, but because VBS system descriptions are a real genre and this one is clear and honest.","headline":"A short VBS system description that makes plausible but unmeasured claims about video retrieval improvements; useful for the competition community, but it doesn't establish any of its central results.","tokens_in":4957,"tokens_out":1857,"would_cite":false,"duration_ms":18744,"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":"The paper claims that diveXplore 6.0 cures missed known-item searches by sampling one keyframe per second instead of one per ten seconds, and improves map retrieval, adds OCR and speech-to-text, and sharpens temporal search.","keywords":["interactive video search","known-item search","shot segmentation","keyframe sampling","temporal search","optical character recognition","speech-to-text","video retrieval"],"falsifier":"Take the same corpus and run the same set of known-item search tasks under diveXplore 5.0's ten-second segmentation and 6.0's one-second segmentation, keeping the feature set, ranking, and interface fixed; if short-target success rates do not rise when the sampling interval shrinks, the claim fails. A complementary check: for each previously missed short target, verify whether any one-second keyframe actually falls inside the target's true interval and is retrievable by the target's distinguishing features.","tokens_in":4246,"feed_emoji":"🎬","tokens_out":7856,"duration_ms":75949,"temperature":0.7,"pith_summary":"diveXplore 6.0 is an interactive video retrieval system refined for a live search competition. The paper's central claim is that the version 5.0 system failed known-item search tasks because its uniform ten-second shot segmentation skipped over very short target segments, and that sampling keyframes every second instead retrieves those targets. It also claims that map-based retrieval should move beyond simple feature-frequency matching, that OCR and speech-to-text are useful additions, and that temporal search needs to combine full-segment events with single-keyframe features. A sympathetic reader would care because these are concrete, mechanism-level choices: if the granularity change works, short and temporally specific video queries become solvable without changing the deep-feature vocabulary.","feed_headline":"Video search drops keyframe sampling from 10s to 1s","feed_subtitle":"Shorter shot segmentation should surface very short target clips that a 10-second grid misses.","key_machinery":"The carrier of the argument is the shot segmentation: the procedure that uniformly divides every source video into consecutive scenes and assigns a representative keyframe to each. In version 5.0 this grid was ten seconds long; in version 6.0 it is one second. That grid determines which units can be retrieved, which keyframes have features extracted, and how map search aggregates feature frequencies across a video. Reducing the interval is the load-bearing change, with alternative map matching, OCR, speech-to-text, and temporal search as supporting additions.","core_discovery":"On the paper's own terms, the central discovery is that retrieval granularity, not feature quality, is the bottleneck for finding short moments in a large video corpus. With the old ten-second segmentation, a target segment that lasts a few seconds has no keyframe of its own and must be reached through neighboring shots, which often means it is never found. Version 6.0 therefore samples one keyframe per second, making every second a potential retrieval unit, and adds OCR and speech-to-text features that let users search visible text and spoken words. The paper further asserts that map search should not rely solely on how often a feature appears in a video's shots, and that temporal-context s","pith_inferences":["The paper's rationale implies a sharper test than it runs: the one-second interval should help exactly the short-segment known-item searches that failed under ten-second sampling, and provide little or no gain for targets much longer than a second. That differential is what would confirm the mechanism.","If granularity is the real cause, a cheaper design could first retrieve coarse shots and then re-sample only the candidate videos at one-second resolution, avoiding a tenfold index increase.","The same finer sampling should also help ad-hoc video search for actions with short temporal extent, though the paper only frames the improvement for known-item search."],"forward_implications":["Known-item search tasks aimed at segments shorter than roughly ten seconds should become directly retrievable, because the target itself now appears as a keyframe instead of being buried inside a ten-second shot.","The feature index grows about tenfold under one-second sampling, so the system trades storage and query cost for finer recall.","Adding OCR and speech-to-text means queries can match text visible on screen and spoken words, covering cases visual concepts cannot distinguish.","Replacing pure feature-frequency map matching with alternative similarity methods should improve map retrieval for queries where a feature appears in many unrelated shots.","Temporal search that combines full-segment action events with single-keyframe features should locate activities with sharper start and end points than event labels alone."],"supporting_citations":[{"why":"The previous version of the same system whose ten-second segmentation and map search are the baseline being replaced.","marker":"[6]"},{"why":"Provides the video corpus that diveXplore indexes and searches in the competition.","marker":"[13]"},{"why":"Describes the companion dataset used for evaluation and task design.","marker":"[1]"},{"why":"Supplies the deep image features used for concept queries and map-similarity search.","marker":"[17]"},{"why":"Provides the object-detection features used in shot and map queries.","marker":"[2]"},{"why":"Supplies the action and event features that need temporal context from full segments.","marker":"[9]"},{"why":"Supplies the scene and place features used as query terms.","marker":"[18]"},{"why":"Provides the OCR capability that lets users search visible text in keyframes.","marker":"[3]"},{"why":"Provides the speech-to-text capability that makes spoken content searchable.","marker":"[10]"}],"fun_headline_variants":["Per-second keyframes catch clips 10s grids miss","1s sampling makes every second searchable","Short clips found: diveXplore cuts keyframe interval to 1s","From 10s to 1s: finding fleeting video moments"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The argument assumes that the reason some target segments are never found is that the ten-second shot grid skips over short moments, not that the feature vocabulary, ranking, or query interface hides them; if the misses have another cause, one-second sampling will not improve search results.","fun_headline_variants_meta":{"raw":{"variants":["Per-second keyframes catch clips 10s grids miss","1s sampling makes every second searchable","Short clips found: diveXplore cuts keyframe interval to 1s","From 10s to 1s: finding fleeting video moments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000401,"raw_usage":{"total_tokens":1877,"prompt_tokens":642,"completion_tokens":1235,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":386,"completion_tokens_details":{"reasoning_tokens":1164}},"tokens_in":386,"tokens_out":1235,"duration_ms":9868,"temperature":1.0,"reasoning_tokens":1164,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:52:49.252390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same corpus and run the same set of known-item search tasks under diveXplore 5.0's ten-second segmentation and 6.0's one-second segmentation, keeping the feature set, ranking, and interface fixed; if short-target success rates do not rise when the sampling interval shrinks, the claim fails. A complementary check: for each previously missed short target, verify whether any one-second keyframe actually falls inside the target's true interval and is retrievable by the target's distinguishing features.","supporting_citations":[{"cited_title":"In: Intl","cited_arxiv_id":null,"evidence_quote":"The previous version of the same system whose ten-second segmentation and map search are the baseline being replaced."},{"cited_title":"Insights on the V3C2 Dataset","cited_arxiv_id":"2105.01475","evidence_quote":"Provides the video corpus that diveXplore indexes and searches in the competition."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Describes the companion dataset used for evaluation and task design."},{"cited_title":"IEEE Trans- actions on Pattern Analysis and Machine Intelligence 42(2), 502–508 (2020), https://doi.org/10.1109/TPAMI.2019.2901464","cited_arxiv_id":null,"evidence_quote":"Supplies the action and event features that need temporal context from full segments."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence 40(6), 1452–1464 (2018), https://doi.org/10.1109/TPAMI.2017.2723009","cited_arxiv_id":null,"evidence_quote":"Supplies the scene and place features used as query terms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OCR capability that lets users search visible text in keyframes."},{"cited_title":"In: IEEE 2011 workshop on automatic speech recognition and understand- ing","cited_arxiv_id":null,"evidence_quote":"Provides the speech-to-text capability that makes spoken content searchable."}],"review_version":1}