{"id":"7f81a6ba-489a-43bb-aeac-1b16ac63fc2e","arxiv_id":"2501.05072","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SPR, a segment-proposal-ranking pipeline with offline indexing and approximate nearest neighbor retrieval, achieves state-of-the-art NDCG on TVR-Ranking while processing a query in under one second.","lead":"A three-stage search framework cuts videos into indexed segments, retrieves the most relevant ones with fast nearest-neighbor lookup, and refines them into ranked moments, answering queries in under a second on 20,000 videos. It reports the best published results on the TVR-Ranking benchmark, though the benchmark and baselines come from the same research group.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on a comparison with same-group baselines copied from [31]; without independent reproduction or error bars, the reported margin over ReLoCLNet is not fully verified.","rationale":"The reader's weakest_assumption is the segment-level relevance premise: that top-200 segment retrieval followed by merging recovers ground-truth moments. The paper actually provides substantial evidence against this being the limiting failure point on TVR-Ranking. Table 4 reports an upper bound of 0.8842 NDCG@10 IoU>=0.7 with tau_C=8, meaning the retrieval-plus-proposal stage can in principle support very high performance; Table 6 shows that increasing the number of retrieved segments beyond 200 does not recover new target moments. Thus, on the evaluated dataset, the segment premise is quantified and appears to hold. The more load-bearing concern is the epistemic foundation of the SOTA claim itself: all baselines come from the same group's dataset paper, the proposed model's best configuration is selected on the same validation split without error bars, and no independent implementation is run. This concern does not change the reader's CONDITIONAL verdict, because the reader already flagged self-referential evaluation and missing error bars as weaknesses. However, it is not the same as the reader's stated weakest_assumption, hence the 'disagree' on that field. The paper is a solid engineering contribution, but the abstract's 'state-of-the-art' assertion should be read as provisional until the comparison is independently and statistically verified.","tokens_in":22886,"tokens_out":10910,"duration_ms":113310,"concrete_test":"Run the official code releases of ReLoCLNet, CONQUER, and XML on TVR-Ranking using the same pseudo-training set (top-40 moments, overlap threshold 0.3), the same optimizer settings, and the same NDCG@K/IoU evaluation script used in [31]; also run SPRReLo-L with five random seeds and report mean and standard deviation. If ReLoCLNet reproduces 0.4353 NDCG@10 IoU>=0.3 and SPRReLo's mean minus two standard deviations remains above that baseline, the SOTA claim is verified. If the reproduced baseline is substantially higher or the margin falls within noise, the abstract should be softened to 'competitive' or 'best among compared configurations.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the state-of-the-art result in Table 1, where SPRReLo reaches NDCG@10 IoU>=0.3 of 0.5509 on test versus 0.4353 for ReLoCLNet. For this claim to hold, the baselines must be strong and fairly configured. The paper states in Section 4.3 that all baseline results are sourced from [31], the TVR-Ranking dataset paper, which shares two authors with this work. No independent rerun of XML, CONQUER, or ReLoCLNet is reported. The SPRReLo variant uses the same feature pipeline and code lineage as the ReLoCLNet baseline, and the best training configuration is selected on the validation set with no error bars or multiple-seed analysis. The load-bearing risk is therefore that the large margin reflects a weak or unfavorably tuned baseline rather than the SPR framework itself. This is a verification gap, not an allegation of misconduct; but until the comparison is independently reproduced, the abstract's 'state-of-the-art' assertion is not fully supported. The framework's internal design is coherent, and the reported upper bounds in Table 4 suggest the retrieval stage is not the main bottleneck, but the headline claim depends on a comparison that has not yet been independently validated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPR, a three-stage framework for Ranked Video Moment Retrieval (RVMR): offline indexing of fixed-length video segments, online top-k segment retrieval with approximate nearest neighbor search, rule-based merging of retrieved segments into coarse proposals, and a refinement/re-ranking module that adjusts timestamps and re-ranks proposals. The framework is evaluated on the TVR-Ranking dataset, with two instantiations (SPRReLo and SPRCLIP) reporting improvements over published baselines, together with efficiency and scalability measurements and an upper-bound analysis of the segment-retrieval stage. The central claims are that SPR reaches state-of-the-art performance on TVR-Ranking while processing a query in about one second and scaling to nearly a million indexed segments.","tokens_in":23078,"tokens_out":5687,"duration_ms":55258,"significance":"If the empirical claims hold, the paper makes a useful contribution: it addresses a realistic retrieval setting where multiple moments can match a query, and it does so with a modular pipeline that is substantially faster than end-to-end moment-localization models. The strengths of the paper are its clean decomposition into three stages, the inclusion of an explicit upper-bound analysis (Table 4) that separates the retrieval bottleneck from the refinement bottleneck, the efficiency and scalability measurements (Tables 2 and 3), detailed ablations of the training strategies (Table 5), and the release of code. The main reservations concern verification: the SOTA comparison relies entirely on baseline numbers taken from the same group's dataset paper, and no variance or significance information is reported, so the margin claimed in the abstract is not yet fully established.","major_comments":[{"comment":"The headline state-of-the-art claim rests entirely on baseline numbers sourced from [31], a paper that shares authors with this submission. Since no baseline is re-run and no variance or significance information is reported, the large margin (e.g., 0.5509 vs 0.4353 for NDCG@10 with IoU>=0.3) cannot be distinguished from favorable configuration or evaluation differences. Please provide at least one independent re-run of the baselines, or report multiple seeds with means and standard deviations for all methods, and explicitly verify that the pseudo-training configuration is identical for all models.","section":"4.3, Table 1"},{"comment":"The abstract claims significant reductions in computational cost and processing time, but Table 2 reports only absolute timings for the SPR stages. There is no comparison with the inference time of XML, CONQUER, or ReLoCLNet under the same hardware. Without such a comparison, the efficiency advantage over existing RVMR methods is not established. Please add wall-clock timings for the baselines or restrict the claim to absolute throughput.","section":"4.4, Table 2"},{"comment":"The training procedure for the refinement and re-ranking module is underspecified. The text says it follows [59], but [59] is trained on full-video ground-truth moments, whereas here the module consumes coarse proposals produced by the first two stages. The paper does not define how training proposals are generated from the pseudo-training set, how positive and negative proposals are sampled within a batch, or how the strong/weak positive selection described in Section 4.7 is performed at training time. This information is needed to reproduce Table 5 and to audit the gains attributed to the added hinge losses.","section":"3.3.2 and 4.2"},{"comment":"The framework's key premise is that every relevant moment is covered by a small set of individually query-relevant segments. On TVR-Ranking this premise is empirically well-supported by the upper-bound analysis (UB NDCG@10 with IoU>=0.7 reaches 0.8842 with tau_C=8), so I do not regard it as a fatal flaw. However, the paper should explicitly acknowledge that for queries whose relevance is only visible at the multi-segment level, top-k segment retrieval cannot recover the moment, and it should either report failure cases or soften the claim that the framework handles videos of any length in general.","section":"3.1, Table 4"}],"minor_comments":[{"comment":"There are several typos: 'additinal' in Table 3's caption, 'an text query' in Section 1, and repeated 'e.g.,,' in the introduction and Figure 1 caption.","section":"Throughout"},{"comment":"The formatting of Table 2 is confusing: the SPIVF and SPIVFPQ rows appear to have a different number of columns than the other rows, making the per-stage time columns ambiguous.","section":"4.4, Table 2"},{"comment":"In Table 5, the row labeled 'SP' has a dash in the Model group column; it would be clearer to label it 'SP (coarse, no refinement)' to avoid confusion with the SPR variants.","section":"5, Table 5"},{"comment":"The caption of Table 1 says 'SP, SPR CLIP and SPR ReLo' but the table uses underscores; please make the notation consistent across text, tables, and figures.","section":"1, Table 1"},{"comment":"The caption of Table 3 contains a fragment 'T: TVR-Ranking validation dataset additinal videos from C: Charades'; please rephrase it as a complete sentence and fix the typo.","section":"4.5, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The author overlap between this submission and [31], the TVR-Ranking dataset paper, makes the SOTA comparison unusually self-referential. I am not alleging misconduct, but an independent evaluation of at least one baseline would substantially strengthen the headline claim. The manuscript is also labeled 'Research in progress' and formatted as a draft; the editor may wish to check whether the submission meets the journal's guidelines for completeness before sending it for another round."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing you should know about this paper is that it is the first practical, modular pipeline for ranked video moment retrieval, and it is mostly honest about what it does and doesn't prove. The SPR decomposition — segment retrieval, rule-based proposal merging, refinement/reranking — is a clean integration of known components: CLIP features, MIL-NCE contrastive loss, Faiss ANN index, and a ReLoCLNet-style refiner. That is real novelty, though it is engineering novelty rather than a new learning paradigm. Indexing fixed 4-second segments offline is what lets the framework claim to handle videos of any length, because online search only touches a few hundred retrieved segments.\n\nCredit where earned: the efficiency numbers are believable (sub-second query over 20K videos), the scalability test with added unrelated videos is a useful sanity check, and the upper-bound analysis is an unusually honest way to locate the bottleneck — it shows the retrieval stage is not the main ceiling. The ablations are thorough, the code is released, and the baselines are matched on pseudo-training set size.\n\nThe soft spots are real but not fatal. The headline SOTA claim is not independently verified: all three baselines come from [31], the dataset paper that shares three authors with this work. No one re-ran XML, CONQUER, or ReLoCLNet, and there are no error bars or multiple-seed runs. The margin over ReLoCLNet is large and consistent, so I think the SPR pipeline is likely better, but the self-referential comparison leaves the claim under-supported. Also, the abstract's hour-long video claim is overreach: TVR-Ranking videos average 76 seconds. The scalability experiment adds more videos, not longer ones; hour-long videos are never tested. One minor point: the best training objective per variant is chosen on validation without variance, so there is some optimistic-selection risk, but it looks modest.\n\nOverall, this paper is for researchers in video moment retrieval, dense retrieval, and practical video search. It deserves a serious referee, and a revision that independently reproduces the baselines or adds error bars plus a genuine long-video evaluation would make the central claim solid. I would send it to peer review, and I would not take the SOTA margin at face value until that verification is done.","headline":"SPR is a clean, honest framework for ranked video moment retrieval, but the SOTA claim needs independent baselines and a real long-video test before I trust the margin.","tokens_in":23672,"tokens_out":3258,"would_cite":true,"duration_ms":29892,"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":"Video moment search drops to under one second per query","keywords":["ranked video moment retrieval","video moment search","segment-proposal-ranking","dense retrieval","approximate nearest neighbor","contrastive learning","TVR-Ranking","video corpus moment retrieval"],"falsifier":"Take a set of queries deliberately constructed so that relevance is only apparent across several segments (for example, 'the scene where she realizes the clue and then acts on it' with the ground truth spanning both parts), retrieve the top-200 segments, and compute how many ground-truth moments are contained in the resulting proposals; if that coverage falls far below the paper's reported upper bound of NDCG@10 IoU≥0.7 = 0.8842, then the assumption that segment-level retrieval suffices is false.","tokens_in":22603,"feed_emoji":"🎬","tokens_out":6050,"duration_ms":53742,"temperature":0.7,"pith_summary":"This paper claims that ranked video moment search can be reduced to a three-stage pipeline: retrieve the most relevant 4-second segments from an offline index, merge adjacent retrieved segments into coarse proposals, then refine the timestamps and re-rank the few dozen proposals. The authors argue that this decomposition removes the 'single perfect match' assumption of earlier video moment retrieval, handles hour-long videos by construction, and keeps online cost near real time even on corpora of hundreds of thousands of segments. On the TVR-Ranking benchmark their SPRReLo instantiation reports NDCG@10 at IoU≥0.3 of 0.5509 on test, against 0.4353 for the ReLoCLNet baseline, with total query processing of roughly 0.7–1 second, and the coarse proposals alone already exceed the baseline at the IoU≥0.3 threshold. The flexible stage separation means any of the three components can be upgraded independently, which is the main practical payoff.","feed_headline":"Video moment search drops to under one second per query","feed_subtitle":"SPR's three-stage segment-proposal-ranking beats prior baselines on TVR-Ranking while staying scalable.","key_machinery":"The load-bearing object is the ordered trio of stages, Segment-Proposal-Ranking: (1) segment retrieval with a shared embedding space and cosine similarity, (2) rule-based proposal generation that merges temporally adjacent segments from the same video, and (3) a refinement and re-ranking head that outputs a matching score φ(Q,M) and predicted start/end timestamps (t_s,t_e). The mechanism that carries the efficiency claim is the fixed segment length of 4 seconds with offline precomputation and approximate nearest neighbor search, so online search cost depends on the number of retrieved segments (k=200), not on video length or corpus size. Training uses the MIL-NCE contrastive loss to align queries with the possibly multiple relevant segments in each batch, and the refinement module is trained with the same losses as ReLoCLNet—video contrastive and hinge loss for the score, frame contrastive and localization loss for boundaries—with an added 8-second context padding to raise the coverage ceiling.","core_discovery":"The central claim is that a ranked list of video moments for a text query can be obtained by first treating each fixed-length video segment as an independent retrieval unit. Segments are embedded with CLIP, a frozen vision-language model, projected into a shared query–segment space learned with the MIL-NCE contrastive objective, and stored in a Faiss ANN index built offline; at query time only the top-200 segments are fetched, so no matter how long the video or how large the corpus, online work is bounded by 200 segments plus the proposals they form. Adjacent segments from the same video are merged into coarse proposals, and a light refinement module—instantiated either with CLIP features or with the ReLoCLNet encoder trained with contrastive and hinge losses—adjusts boundaries and re-scores moments. The paper's evidence is that this two-level (coarse then fine) design beats all prior RVMR baselines on TVR-Ranking by a substantial margin, cuts average per-query time to under one second across 19,614 videos, and degrades only slightly when the corpus is tripled with irrelevant videos.","pith_inferences":["Beyond the paper: the same segment-proposal-rank recipe should transfer to other sub-second temporal units and other modalities (audio, subtitles) as long as the segment embedding captures query-relevant content; the paper only evaluates visual features, mostly CLIP frames and I3D/RoBERTa subtitle streams in one instantiation.","Beyond the paper: the framework's dependence on per-segment relevance means it is likely to miss 'emergent' moments whose relevance is defined by a multi-segment relationship, such as cause-effect chains across scenes; a natural test is whether top-200 segment recall on such queries falls well below the reported upper bound of NDCG@10 IoU≥0.7 = 0.8842.","Beyond the paper: the reported collapse of the IVFPQ index (NDCG dropping from roughly 0.44 to 0.09) suggests that product quantization is unsafe for this embedding distribution; a testable extension is to add a two-stage re-rank over an IVF shortlist to recover accuracy at comparable speed.","Beyond the paper: the pseudo-training set built with SimCSE similarity is a weak supervision signal, and a dataset with manual relevance grades across moments would likely close part of the gap between current results and the practical upper bound."],"forward_implications":["If the framework's claims hold, ranked moment search becomes practical at the scale of tens of thousands of videos, with per-query latency of about one second on a single GPU.","Because the three stages are independent, improvements in embedding models, ANN indexes, or refinement architectures can be dropped in without re-architecting the rest of the system.","The upper-bound analysis shows that even a perfect refinement module cannot recover moments that fall outside the top-200 retrieved segments, so improving segment recall is the direct path to raising the performance ceiling.","The scalability experiments indicate that adding large numbers of unrelated videos does not hurt ranking quality when using Flat or IVF indexes, supporting deployment on growing corpora."],"supporting_citations":[{"why":"Defines the RVMR task, supplies the TVR-Ranking dataset with relevance annotations and the NDCG@K, IoU-threshold metric, and provides the baseline results the framework must beat.","marker":"[31]"},{"why":"Supplies the frozen CLIP vision-language model used to extract frame and query features in the main instantiation of segment retrieval and refinement.","marker":"[44]"},{"why":"ReLoCLNet serves as the strongest baseline and as the backbone of one refinement instantiation, including the contrastive and hinge training losses reused by this paper.","marker":"[59]"},{"why":"Provides the Faiss library used for exact (Flat) and approximate (IVF, IVFPQ) nearest neighbor search over the precomputed segment embeddings.","marker":"[11]"},{"why":"Supplies the MIL-NCE contrastive loss used to align a query with its multiple relevant segments during projector training.","marker":"[39]"},{"why":"Provides the raw TVR videos and moment annotations from which TVR-Ranking is built, plus the I3D and BERT features used in the ReLo-based refinement instantiation.","marker":"[28]"},{"why":"Supplies the SimCSE similarity used to assign pseudo-training moments for each query in the TVR-Ranking training set.","marker":"[15]"}],"fun_headline_variants":["SPR finds video moments in under one second","Video moment search: one second, any video length","Scalable video moment search with SPR's 3-stage design","Ranked video moment retrieval made fast and flexible","SPR cuts video moment search to sub-second speeds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that any ground-truth moment is well covered by a handful of individually query-relevant 4-second segments, so that searching the top-200 segments never discards the truth before refinement begins.","fun_headline_variants_meta":{"raw":{"variants":["SPR finds video moments in under one second","Video moment search: one second, any video length","Scalable video moment search with SPR's 3-stage design","Ranked video moment retrieval made fast and flexible","SPR cuts video moment search to sub-second speeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1343,"prompt_tokens":999,"completion_tokens":344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":268}},"tokens_in":615,"tokens_out":344,"duration_ms":4200,"temperature":1.0,"reasoning_tokens":268,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:19:18.632974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of queries deliberately constructed so that relevance is only apparent across several segments (for example, 'the scene where she realizes the clue and then acts on it' with the ground truth spanning both parts), retrieve the top-200 segments, and compute how many ground-truth moments are contained in the resulting proposals; if that coverage falls far below the paper's reported upper bound of NDCG@10 IoU≥0.7 = 0.8842, then the assumption that segment-level retrieval suffices is false.","supporting_citations":[{"cited_title":"TVR-Ranking: A Dataset for Ranked Video Moment Retrieval with Imprecise Queries","cited_arxiv_id":"2407.06597","evidence_quote":"Defines the RVMR task, supplies the TVR-Ranking dataset with relevance annotations and the NDCG@K, IoU-threshold metric, and provides the baseline results the framework must beat."}],"review_version":1}