{"id":"60393b15-bdea-484c-9d4b-9c7a9d2d5bf7","arxiv_id":"2606.31421","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"TemporalLens diagnoses temporal reliance in single-stage video detectors via controlled perturbations, while YOLO-3D shows temporal preservation yields +3.7 pp mAP@50 gains.","lead":"The paper introduces TemporalLens, a diagnostic framework using perturbations like frame removal and shuffling to test whether video detectors rely on temporal context across frames or just one informative frame. It also presents YOLO-3D, claiming that preserving temporal depth in the backbone drives most performance gains in real-time spatiotemporal detection.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Perturbations may not isolate temporal dependence without spatial or internal confounds","rationale":"The reader's weakest assumption directly identifies the load-bearing risk for both the diagnostic framework and the architecture claim. Full-text access does not remove this uncertainty; the reported experiments would need the suggested isolation check to strengthen attribution. No other internal inconsistency appears load-bearing from the given claims.","tokens_in":1724,"tokens_out":325,"duration_ms":17354,"concrete_test":"Re-run the frame-removal and temporal-shuffle experiments on a frozen single-frame YOLOv8 baseline (identical spatial backbone, no temporal input) and compare mAP drop magnitudes and prediction recovery patterns to those reported for the stacked 2D and YOLO-3D models; if the single-frame model exhibits comparable collapse/recovery, the perturbations do not isolate temporal effects.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The dominant-driver claim (+3.7 pp mAP@50 from preserving temporal depth) and the diagnostic conclusions both rest on TemporalLens perturbations cleanly separating temporal reliance from spatial feature quality or model-internal changes. Frame removal, shuffling, and redundancy injection alter the spatiotemporal input statistics; for stacked-frame 2D backbones this could degrade spatial feature extractors or change activation distributions in ways unrelated to temporal reasoning, while 3D models might respond differently due to architectural inductive biases rather than genuine cross-frame reasoning. Without explicit controls (e.g., measuring spatial-only performance under identical perturbations or verifying feature-map statistics), the observed behavioral gap cannot be attributed solely to temporal dependence.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces TemporalLens, a model-agnostic diagnostic framework that applies controlled perturbations (frame removal, temporal shuffling, redundancy injection, resolution degradation) to probe whether single-stage video detectors use temporal context or rely on single frames. Applied to stacked-frame 2D detectors and the proposed YOLO-3D architecture, it reports behavioral differences (stacked 2D collapse without target frame; 3D recovers from prior frames) invisible to mAP, and claims that preserving temporal depth through the backbone is the dominant driver (+3.7 pp mAP@50 at 32 frames averaged across scales).","tokens_in":1845,"tokens_out":427,"duration_ms":36944,"significance":"If the perturbations are validated to isolate temporal dependence, the work supplies a concrete diagnostic for temporal reasoning in video detectors and a modular, real-time architecture whose key insight is simple depth preservation rather than complex temporal processing. This addresses a gap between standard mAP and actual temporal use, with potential to guide efficient spatiotemporal detector design.","major_comments":[{"comment":"Abstract: the central claims that TemporalLens reveals genuine temporal reliance (stacked 2D collapse vs. 3D recovery) and that temporal-depth preservation drives +3.7 pp mAP@50 both rest on the unverified assumption that the listed perturbations isolate temporal dependence. Frame removal, shuffling, and redundancy injection alter spatiotemporal statistics; without explicit controls (spatial-only performance under the same perturbations, or feature-map statistic verification), the observed gap could arise from differential spatial-feature degradation or 3D inductive biases rather than cross-frame reasoning.","section":"Abstract"},{"comment":"Abstract: the performance number (+3.7 pp mAP@50 at 32 frames averaged across scales) and the behavioral-difference claims are presented without error bars, dataset specifications, frame-count details, or any description of how the perturbations were validated to cleanly measure temporal use.","section":"Abstract"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which identify key areas where additional validation and reporting details will strengthen the manuscript. We address each point below and will incorporate the suggested improvements in the revision.","responses":[{"response":"We agree that the manuscript would benefit from explicit controls to further isolate temporal effects. In the revision we will add experiments applying the same perturbation types in a purely spatial manner (e.g., spatial shuffling or resolution degradation applied identically across frames) and will include feature-map statistic comparisons (mean activation, variance, and correlation across layers) between the stacked-2D and YOLO-3D models under each condition. These additions will directly address the possibility of spatial-feature degradation or 3D inductive biases as alternative explanations.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claims that TemporalLens reveals genuine temporal reliance (stacked 2D collapse vs. 3D recovery) and that temporal-depth preservation drives +3.7 pp mAP@50 both rest on the unverified assumption that the listed perturbations isolate temporal dependence. Frame removal, shuffling, and redundancy injection alter spatiotemporal statistics; without explicit controls (spatial-only performance under the same perturbations, or feature-map statistic verification), the observed gap could arise from differential spatial-feature degradation or 3D inductive biases rather than cross-frame reasoning."},{"response":"We acknowledge that the abstract and main text currently omit these reporting elements. The revised manuscript will expand the abstract and methods sections to report error bars (standard deviation across three independent runs), name the exact datasets and splits used, specify the precise frame counts and averaging procedure, and add a subsection describing perturbation validation (including the spatial-control experiments noted above).","revision_made":"yes","referee_comment":"[Abstract] Abstract: the performance number (+3.7 pp mAP@50 at 32 frames averaged across scales) and the behavioral-difference claims are presented without error bars, dataset specifications, frame-count details, or any description of how the perturbations were validated to cleanly measure temporal use."}],"tokens_in":1414,"tokens_out":450,"duration_ms":26548,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main contribution is TemporalLens, a diagnostic that applies frame removal, temporal shuffling, redundancy injection, and resolution degradation to check temporal dependence in single-stage video detectors. On stacked 2D models it shows collapse when the target frame is gone, while their YOLO-3D recovers from earlier frames. They also report that simply preserving temporal depth through the backbone accounts for the bulk of the improvement, +3.7 pp mAP@50 at 32 frames averaged across scales.\n\nThis is useful because mAP alone does not distinguish models that reason over time from those that latch onto one informative frame, and the perturbations turn that distinction into something observable for time-critical deployments.\n\nThe soft spot is whether the perturbations cleanly isolate temporal reliance. Frame removal and shuffling change input statistics that could degrade spatial feature extraction or activation patterns differently in 2D versus 3D architectures, and the abstract gives no controls such as spatial-only performance under the same changes or checks on feature-map statistics. Without those, the behavioral gap might partly reflect architectural inductive biases rather than genuine cross-frame reasoning. The performance delta also lacks error bars or dataset specifics here.\n\nThis is aimed at people designing or tuning real-time video detectors who need to verify actual temporal use. It has a concrete method and a clear result worth referee time, even if the isolation claim needs tighter validation.","headline":"The paper gives a practical way to test if video detectors actually use time across frames and finds that keeping temporal depth in the backbone drives most gains in their YOLO-3D variant.","tokens_in":2324,"tokens_out":358,"would_cite":false,"duration_ms":26119,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Preserving temporal depth through the backbone accounts for most accuracy gains in single-stage video detectors.","keywords":["video object detection","temporal reasoning","single-stage detectors","spatiotemporal models","YOLO","diagnostics","real-time detection","frame perturbations"],"falsifier":"An experiment in which a spatiotemporal model achieves identical mAP when every frame except one is removed as when the full sequence is supplied would falsify the claim that temporal preservation drives the gains.","tokens_in":2635,"feed_emoji":"📹","tokens_out":681,"duration_ms":19187,"temperature":0.7,"pith_summary":"Many single-stage video detectors appear to improve with more frames but actually rely on whichever single frame happens to be clearest. The work introduces a diagnostic suite of frame removals, shuffles, and redundancy injections that exposes this difference: stacked 2D models lose performance when the target frame disappears, while models that keep full temporal depth can still recover earlier information. A modular architecture built on YOLOv8 shows that simply maintaining the original time dimension through the backbone produces the largest lift, +3.7 points mAP@50 at 32 frames averaged across scales. These two pieces together convert the question of whether a detector reasons over time into a concrete, testable property rather than an assumption hidden inside overall accuracy numbers.","feed_headline":"Preserving temporal depth drives most video detector gains","feed_subtitle":"A diagnostic suite shows stacked 2D models ignore time while keeping full frame depth in the backbone adds 3.7 mAP points at 32 frames.","key_machinery":"TemporalLens, a model-agnostic set of controlled perturbations including frame removal, temporal shuffling, redundancy injection, and resolution degradation that measures whether predictions depend on information across multiple frames.","core_discovery":"The paper establishes that temporal preservation, rather than specialized temporal processing operations, is the dominant factor behind performance in spatiotemporal single-stage detectors. Controlled perturbations demonstrate that models without preserved temporal depth collapse when the target frame is removed, whereas models that retain depth recover predictions from prior frames. The YOLO-3D design isolates this effect by keeping the full temporal dimension intact through the backbone and records the reported accuracy improvement without requiring additional fusion mechanisms.","pith_inferences":["Designers could replace complex 3D modules with simple depth-preserving backbones in real-time settings.","The same perturbation tests could be applied to action recognition or video segmentation to check temporal reliance.","Training schedules might be adjusted to penalize single-frame shortcuts once the diagnostics are in place."],"forward_implications":["Stacked 2D detectors fail the removal test and therefore do not reason over time.","Spatiotemporal models that keep temporal depth recover usable predictions from earlier frames after the target frame is deleted.","The largest accuracy increase comes from maintaining temporal depth rather than adding temporal operators.","The diagnostic turns the presence of temporal reasoning into a measurable behavioral signature."],"fun_headline_variants":["Temporal preservation trumps processing in detectors","Stacked models fail without target frame removal tests","Depth retention alone adds 3.7 mAP at 32 frames","YOLO-3D shows temporal depth as dominant factor","Perturbations expose collapse without preserved depth"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The chosen perturbations change only the availability of temporal context and do not alter spatial feature processing or internal model behavior in ways that would produce the observed differences by accident.","fun_headline_variants_meta":{"raw":{"variants":["Temporal preservation trumps processing in detectors","Stacked models fail without target frame removal tests","Depth retention alone adds 3.7 mAP at 32 frames","YOLO-3D shows temporal depth as dominant factor","Perturbations expose collapse without preserved depth"]},"model":"grok-4.3","cost_usd":0.006456,"raw_usage":{"total_tokens":3031,"prompt_tokens":682,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":64562000,"prompt_tokens_details":{"text_tokens":682,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2285,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":682,"tokens_out":64,"duration_ms":24952,"temperature":1.0,"reasoning_tokens":2285,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-01T05:28:38.584137+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An experiment in which a spatiotemporal model achieves identical mAP when every frame except one is removed as when the full sequence is supplied would falsify the claim that temporal preservation drives the gains.","supporting_citations":[],"review_version":1}