{"id":"4647a6fa-4a63-488a-b016-37006a83ffbf","arxiv_id":"2607.13421","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A coarse-to-fine video-grounding framework improves temporal boundary accuracy by densely re-examining frames around coarse start/end predictions, achieving SOTA on HC-STVGv1/v2 and VidSTG.","lead":"This paper presents ScanFocus, a two-stage system for locating an object in a video from a text description: first it scans a sparse sample of the whole video, then it zooms in on the predicted start and end moments and re-examines dense frames around them. The authors report state-of-the-art accuracy on three standard video-grounding benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Refinement recoverable range is only ±1.75 coarse sampling intervals; paper never measures whether GT boundaries fall inside the dense windows.","rationale":"The reader's weakest assumption approximately matches my own read: the refinement stage's ability to recover high-frequency boundary cues is contingent on the coarse interval being close enough that the true boundaries fall inside the small dense windows. My analysis sharpens the point by giving the exact recoverable range (1.75 coarse sampling intervals for Nw=8, fr=2fc), and by noting that neither the final metrics nor the oracle experiment can validate this prerequisite. This is not a claim of fraud or sloppiness; it is a missing measurement that is directly load-bearing for the central 'recovery' narrative. The backbone confound noted by the reader is also real, but the coarse-window coverage issue is more central to the proposed mechanism itself. Since the current evidence is suggestive but incomplete, I would keep the CONDITIONAL verdict and ask for the coverage/recall measurement as a condition of acceptance.","tokens_in":14959,"tokens_out":6437,"duration_ms":66418,"concrete_test":"On HC-STVGv1 test, run the frozen coarse model. For each sample, measure coverage: do the GT start/end timestamps lie within the refine windows around predicted tcs/tce, i.e., |t_pred − t_gt| ≤ (Nw−1)/(2fr)? Report the hit rate separately for starts and ends, and breakdown final m_tIoU/vIoU for hit vs miss subsets. If coverage is below ~95%, or if refinement gains vanish on the miss subset, the coarse-to-fine recovery claim fails; if coverage is high, the concern is retired.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 samples Nw=8 dense frames at fr=2·fc inside windows centered on the coarse predictions tcs/tce. The maximum boundary offset the refine stage can absorb is therefore (Nw−1)/(2fr)=7/(4fc)=1.75 coarse frame intervals; any GT boundary farther away is physically outside the sampled support and cannot be recovered, no matter how strong SGTA is. The paper reports only final metrics and the Fig. 2 oracle replaces predicted intervals with GT timestamps, so neither measurement tests this prerequisite. It also does not state whether refine-stage training centers windows on predicted or GT coarse boundaries; a GT-centered training window would make the test-time dependency worse. Without a reported coarse-interval recall / window hit-rate, the 2.8-point tIoU gain (Table 4: 50.9→53.7) and the SOTA claim remain consistent with refinement helping only the subset of samples whose coarse boundaries happen to land close. This is an unvalidated load-bearing assumption, not an internal contradiction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ScanFocus proposes a two-stage coarse-to-fine framework for spatio-temporal video grounding (STVG). A global stage uniformly samples T_c=64 frames at a low rate, fuses BEiT-3 appearance/text features with VideoMAE motion features via a lightweight deformable semantic-motion fusion encoder, and generates coarse spatial tubes and a coarse temporal interval with dual DETR-style decoders. A local refinement stage then densely samples N_w=8 frames at twice the coarse rate around each predicted boundary, applies the Semantic-Guided Temporal Aggregator (SGTA) to model short-term inter-frame dependencies, and regresses refined start/end timestamps. The paper reports state-of-the-art numbers on HC-STVGv1, HC-STVGv2, and VidSTG, an oracle experiment attributing most remaining error to temporal localization, and ablations over components, window size, SGTA design, and fusion mechanism.","tokens_in":15230,"tokens_out":8174,"duration_ms":70724,"significance":"If the reported results are reproducible, ScanFocus is a solid contribution to STVG: it targets a real limitation of low-frame-rate global modeling, provides a clean architecture that decouples global context from local boundary precision, and reports consistent gains across three benchmarks. The oracle analysis in Fig. 2 is a valuable diagnostic, and the component ablations in Tables 4–7 generally support the design choices; the efficiency gain in Table 7 is also a concrete plus. The main risk is that the refinement stage's central assumption—that coarse boundaries fall within ±1.75 coarse sampling intervals—is not validated. Without a hit-rate analysis, the mechanism claimed to recover high-frequency cues is not actually demonstrated.","major_comments":[{"comment":"The refinement windows sample N_w=8 frames at f_r=2·f_c centered on the predicted coarse boundaries. The largest boundary offset that can possibly be recovered is therefore (N_w−1)/(2·f_r)=7/(4·f_c)≈1.75 coarse frame intervals. The paper never reports the distribution of coarse-boundary errors, the fraction of GT boundaries that fall inside the sampled windows, or whether refine-stage training windows are centered on predicted or GT coarse boundaries. Without this, the 2.8-point tIoU improvement in Table 4 (50.9→53.7) and the SOTA claim remain consistent with refinement helping only samples with already-accurate coarse predictions; GT-centered training would additionally create a train/test mismatch. Please report coarse-interval recall / window hit-rate and the exact training window centering.","section":"Sec. 3.4 (Dense Boundary Sampling)"},{"comment":"Tables 4–6 report single runs with no error bars, standard deviations, or significance tests. Some of the claims rest on small differences: Full SGTA vs w/o SG is 0.6 tIoU and vs w/o TA is 0.9 tIoU in Table 6. Moreover, ablations are run at 224×224, while the main comparisons in Tables 1–3 use 384×384 and a different backbone than prior methods, so the apparent SOTA margin may partly reflect the choice of BEiT-3/VideoMAE rather than the coarse-to-fine design. Please report multi-seed statistics and a same-backbone, same-resolution baseline (e.g., the coarse global stage without refinement) in the main results.","section":"Sec. 4.3, Tables 4–6"},{"comment":"Sec. 4.1 states that detailed network configurations and loss weights are provided in the Supplementary Material, but no supplementary document is included with arXiv v1, and the paper also refers to it for the other datasets' oracle experiments (Sec. 4.1, Fig. 2 caption). Without specifying the number of encoder/decoder layers, attention heads, loss weights λ_box/λ_tmp/λ_ref/λ_act, learning rates, and training iterations, the experiments are not reproducible. The refine-stage training protocol is especially important because it determines whether the model sees predicted or GT windows during training.","section":"Sec. 4.1 (Implementation)"},{"comment":"Motion features are defined as F_m ∈ R^{T×N_m×C_m} with T the original video length, while the coarse input consists of only T_c uniformly sampled frames. If VideoMAE is applied to the full T frames, the claimed computational savings from sparse sampling are negated; if it is applied to the sampled frames, the notation is inconsistent and the alignment between F_m and F_a (which has T_c frames) needs to be stated precisely. The same issue affects the complexity analysis in Table 7, which reports GFLOPs but not the temporal resolution used for motion extraction.","section":"Sec. 3.3 (Feature Extraction)"}],"minor_comments":[{"comment":"First sentence: 'motion details that suppressed by' should be 'that are suppressed by'.","section":"Sec. 3.4"},{"comment":"The projections φ_a and φ_t are not dimensionally specified; please give the output dimensions of φ_a(F_ps^a) and φ_t(F_rs^t), or state that all features are projected to a common channel C.","section":"Eq. (5)"},{"comment":"The dense-boundary-sampling region contains many unreadable small labels (e.g., 'cc_132509', '1424', the list of numbers); please enlarge fonts and remove irrelevant tokens.","section":"Fig. 3"},{"comment":"'For brevity, tiou and viou are used to represent m_tIoU and m_vIoU'—define these abbreviations at first use and use consistent capitalization.","section":"Sec. 4.3"},{"comment":"Inputs are resized to 384×384 for main results but ablations use 224×224. Please clarify whether the GFLOPs in Tables 5 and 7 correspond to the 224×224 or 384×384 setting.","section":"Sec. 4.1"},{"comment":"Reference [46] has a typo ('Protege'); reference [55] contains a stray '(7)' in the arXiv ID. Please verify the official bibliographic details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript repeatedly cites a Supplementary Material that is not available in the arXiv v1 submission; the editor should require the supplement and any code release before acceptance. The central claim would be substantially strengthened by a coarse-window hit-rate analysis and by multi-seed error bars. The self-citation to [12] (GC3VG) is related but not circular; however, the authors should be aware that the coarse-to-fine concept is not entirely novel. Also note that the reader's concern about missing same-backbone baselines is partially addressed by the 'Coarse' row in Table 4, but only at 224×224, so the main 384×384 comparisons remain confounded by backbone choice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Decent empirical paper. The new bit is the coarse-to-fine decoupling: a sparse global scan over 64 sampled frames, then dense local windows of 8 frames at 2x the coarse rate around the predicted boundaries, with a Semantic-Guided Temporal Aggregator to model short-term interactions. That directly targets the known problem of downsampling wiping out boundary cues. The ablations are convincing: each component (dense sampling, SGTA, deformable fusion) adds a measurable gain, and the gains are consistent on three benchmarks. The oracle experiment is a nice way to show temporal localization is the bottleneck.\n\nThe softest spot is the unvalidated assumption in the refine stage. With Nw=8 and kr=2, the dense windows cover ±1.75 coarse intervals around the predicted boundaries. If the coarse prediction is off by more than that, the true boundary is not in the input, so refinement cannot fix it. The paper never reports how often the coarse boundaries land close enough, nor a hit-rate for the windows. The oracle in Fig. 2 uses GT timestamps, so it doesn't tell us that. This is a missing diagnostic, not a fatal flaw—the final numbers suggest the coarse stage is usually close—but it should be reported.\n\nOther issues: no error bars or significance tests; implementation details are deferred to a supplementary that isn't there; and there is no same-backbone comparison against a single-stage model with BEiT-3, though the fact that their coarse-only model (50.9) is below TA-STVG (53.0) rather weakens the 'better backbone' explanation. They also cite their own GC3VG without discussing it—a minor positioning gap.\n\nThere's nothing circular or incoherent here. The paper is clearly written, the components are plausible, and the reported gains are consistent. It deserves a serious referee. I would ask the authors to (1) report coarse-window hit-rate and maybe a coarse-error vs. correction plot, (2) add error bars, (3) release code, and (4) explicitly position against GC3VG. For the STVG community this is a useful recipe, and it should be in the literature after those additions.","headline":"A solid, well-ablated coarse-to-fine STVG paper with a real but unmeasured assumption about how close the coarse stage needs to be; worth a serious referee.","tokens_in":15726,"tokens_out":3881,"would_cite":true,"duration_ms":38510,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ScanFocus claims that decoupling spatio-temporal video grounding into a coarse global scan and a dense local boundary-focus stage, driven by the Semantic-Guided Temporal Aggregator, recovers the high-frequency boundary cues that low-rate do","keywords":["spatio-temporal video grounding","coarse-to-fine","boundary refinement","temporal localization","semantic-guided temporal aggregator","dense sampling","multimodal fusion","video grounding"],"falsifier":"Compute the fraction of ground-truth start and end timestamps that lie within the Nw=8 dense windows (sampled at twice the coarse rate) around the model's coarse predictions. If that recall is significantly below 100%—say, under 90%—then the refinement stage is provably unable to recover those missed boundaries, and the gain attributed to SGTA would overstate the recoverable improvement. A complementary experiment: feed the refine stage with coarse intervals artificially perturbed by increasing offsets (0, 2, 4, 6, 8 frames) and measure how quickly precision degrades; the degradation curve rev","tokens_in":14850,"feed_emoji":"🎯","tokens_out":8345,"duration_ms":72178,"temperature":0.7,"pith_summary":"The paper argues that the main obstacle to accurate spatio-temporal video grounding—locating the trajectory of an object described in a sentence, frame by frame—is not spatial detection but temporal boundary ambiguity, caused by the low-rate temporal downsampling that most existing methods use to keep global modeling affordable. ScanFocus tries to prove that decoupling the task into a coarse global scan at a low frame rate and a fine local refinement that densely samples the predicted boundaries recovers the high-frequency cues that downsampling suppresses. At the core is the Semantic-Guided Temporal Aggregator (SGTA), which explicitly models short-term inter-frame dependencies inside each dense boundary window, guided by the referred object's appearance and the language query. The paper reports consistent state-of-the-art results on HC-STVGv1/v2 and VidSTG, with the largest gains on strict metrics like vIoU@0.3 and vIoU@0.5, and an oracle experiment indicating that temporal localization, not spatial localization, is the current bottleneck.","feed_headline":"Coarse scan then dense refine recovers boundary cues, sets STVG record","feed_subtitle":"Dense refinement around coarse boundaries cuts temporal ambiguity, lifting vIoU@0.5 by 5.4 points on HC-STVGv1","key_machinery":"The key machinery is the local boundary-focus stage built around the Semantic-Guided Temporal Aggregator (SGTA). Given a coarse predicted interval, the method densely samples Nw=8 frames at twice the coarse frame rate in two windows centered on the predicted start and end. SGTA then: (1) uses the coarse bounding-box sequence, temporally interpolated to the window length, as a spatial prior for RoI pooling on the appearance features; (2) modulates the motion features by the object-aligned appearance features (Hadamard product) and injects the text features as a bias; and (3) flattens the resulting semantic-guided motion tokens into a sequence and passes them through stacked temporal self-atte","core_discovery":"ScanFocus claims that a coarse-to-fine decomposition of spatio-temporal video grounding—first a sparse global scan that produces a rough temporal interval, then a local boundary-focus stage that densely samples frames around that interval and explicitly models inter-frame dependencies—recovers the high-frequency boundary cues lost to standard low-rate temporal subsampling, and that this recovery is what drives precise start/end timestamp regression. The central evidence is a set of benchmark results: on HC-STVGv1, the method reaches m_tIoU 55.5, m_vIoU 41.8, vIoU@0.3 67.5, and vIoU@0.5 42.2, surpassing the previous best (TA-STVG) by 4.4 and 5.4 points on vIoU@0.3 and vIoU@0.5 respectively; c","pith_inferences":["The paper's fixed window size (Nw=8 at twice the coarse rate) implies a hard reach: any true boundary more than half a window away from the coarse prediction is invisible to the refinement stage. A natural testable consequence is that the method's gains should concentrate on clips where the coarse prediction is already within a few frames of the true boundary; if that recall is low, the reported m","The same coarse-to-fine recipe could be applied in an iterative or cascaded manner (refine the refined boundaries again), or with adaptive window sizes based on coarse confidence, which the paper does not explore.","Because the refinement stage reuses the coarse features and only adds lightweight modules, the approach suggests that the bottleneck in many video-language tasks may be temporal granularity rather than model capacity; a cheaper fix for existing systems might be to add a similar boundary-refinement head rather than retrain the whole model.","If the oracle gap is as large as reported (vIoU@0.3 jumps to 94% with ground-truth timestamps), then even a perfect temporal-refinement module would still leave substantial room before spatial grounding becomes the limiting factor; the paper's framing implies that future work should focus on boundary regression accuracy, perhaps with explicit uncertainty estimation."],"forward_implications":["If the decoupling claim holds, temporal boundary precision in spatio-temporal video grounding can be improved without paying the full cost of dense global processing, since the dense computation is confined to small windows around coarse boundaries.","The oracle experiment implies that the spatial grounding branch is already near-saturated; further gains on these benchmarks are most likely to come from better temporal boundary modeling, not better object detection.","The SGTA design—semantic guidance via RoI pooling on interpolated coarse boxes plus temporal self-attention on flattened tokens—offers a reusable template for other video tasks that need fine temporal boundaries, such as action localization or moment retrieval.","Replacing the heavy MDETR-style tri-modal fusion with a general pre-trained vision-language encoder plus a lightweight deformable fusion reduces computational cost while improving accuracy, suggesting that decoupling perception and fusion can be beneficial beyond this task.","The reported gains are largest on strict temporal-spatial overlap thresholds (vIoU@0.3 and vIoU@0.5), indicating that the method specifically improves boundary sharpness rather than just coarse region agreement."],"fun_headline_variants":["Coarse scan plus dense refine recovers video boundary cues","ScanFocus: coarse-to-fine framework for precise video grounding","Dense sampling near coarse boundaries sharpens grounding","Recovering high-frequency cues with coarse-to-fine design"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The refinement stage can only fix boundaries if the true start and end frames fall inside the dense sampling windows centered on the coarse predictions; if the coarse interval is off by more than about half a window, the refinement never sees the true boundary, and the paper does not measure how often that happens.","fun_headline_variants_meta":{"raw":{"variants":["Coarse scan plus dense refine recovers video boundary cues","ScanFocus: coarse-to-fine framework for precise video grounding","Dense sampling near coarse boundaries sharpens grounding","Recovering high-frequency cues with coarse-to-fine design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000412,"raw_usage":{"total_tokens":2008,"prompt_tokens":821,"completion_tokens":1187,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1120}},"tokens_in":565,"tokens_out":1187,"duration_ms":12159,"temperature":1.0,"reasoning_tokens":1120,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T05:12:38.037095+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the fraction of ground-truth start and end timestamps that lie within the Nw=8 dense windows (sampled at twice the coarse rate) around the model's coarse predictions. If that recall is significantly below 100%—say, under 90%—then the refinement stage is provably unable to recover those missed boundaries, and the gain attributed to SGTA would overstate the recoverable improvement. A complementary experiment: feed the refine stage with coarse intervals artificially perturbed by increasing offsets (0, 2, 4, 6, 8 frames) and measure how quickly precision degrades; the degradation curve rev","supporting_citations":[],"review_version":1}