{"id":"71df6e8b-9a4b-436c-822e-05616eda1e1d","arxiv_id":"1908.07904","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper builds a synthetic motion-blur tracking benchmark, finds light blur helps and heavy blur hurts most trackers, and proposes fine-tuning a DeblurGAN discriminator to selectively deblur frames, improving six of seven tested trackers.","lead":"This paper measures how motion blur changes visual object tracking by building a benchmark of videos with five controlled blur levels. It finds that light blur often helps trackers, heavy blur hurts them, and selectively deblurring only very blurry frames improves six trackers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The §3.1 ground-truth rule averages the L sharp boxes; for moving targets the correct blurred box is their union, so the heavy-blur AUC decline and the central 'blur always hurts' claim may be a GT artifact.","rationale":"The reader's weakest assumption already pointed at the blurred-frame ground truth, and my analysis sharpens it: the issue is not only whether synthetic averaging matches real camera blur, but that the stated rule is geometrically wrong for the blurred object. Averaging boxes underestimates the extent of a moving target, and every heavy-blur comparison in the paper uses these boxes, so the headline 'heavy blur always hurts' and the reported selective-deblurring gains on S16 are not yet established. This is a correctable flaw, and a union-box or human-label re-evaluation would settle it. I considered other candidate concerns: the §4.2 selective-deblurring experiment is oracle-based and thus only illustrative, and the absence of error bars limits confidence, but neither is as load-bearing as the GT rule because the final §5.4 method is not oracle-based and the main claims all flow through the benchmark. I therefore keep the CONDITIONAL verdict and would make acceptance depend on the GT correction or validation.","tokens_in":14087,"tokens_out":6340,"duration_ms":67270,"concrete_test":"Recompute all BVT results with corrected ground truth: for each blurred frame, define the target box as the tight bounding box of the union of the L sharp-frame boxes (per-coordinate min/max of corners), then rerun the 23 trackers on S{2,4,8,16} and recompute AUC, AUC gains, and NRS. If the S1→S16 AUC delta largely disappears or reverses for trackers that follow the smear, the averaged-box GT is the driver of the heavy-blur finding. As a sanity check, human-annotate a random sample of 50 S16 frames and measure IoU between the averaged-box GT and the human labels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §3.1 the blurred frame is \\tilde I_t^L = avg(I_t,...,I_{t+L-1}) and 'the ground truth ... is set as the average of annotations of medium frames.' This is not the box of the time-integrated target. If the target moves horizontally by d during the exposure, the visible smear occupies width w+d, while the averaged box has width w and midpoint center. A tracker that perfectly predicts the smear is scored with IoU ≈ w/(w+d), which decreases with L. Hence the S8/S16 AUC drops and the 'heavy blur always hurts' conclusion may reflect the GT rule, not tracker failure. The NRS metric inherits this bias because the IoU>0.5 success gate from S1 is applied to these undersized boxes. The light-blur benefit on S2/S4 is less affected since d is small, but the selective-deblurring gains on heavily blurred subsets are computed on the same GT and are equally suspect. Frame averaging itself is a standard way to synthesize blur; the specific box-averaging step is the weak link.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper builds a Blurred Video Tracking (BVT) benchmark by temporally averaging frames from the 240 fps NfS dataset at five blur levels (L = 1, 2, 4, 8, 16), with ground-truth boxes defined by averaging the annotations of the averaged frames. Twenty-three trackers are evaluated, leading to the claims that light blur often improves tracking accuracy while heavy blur always hurts it, and that full deblurring helps on heavily blurred videos but hurts on lightly blurred ones. The paper also proposes a blur-robust tracking scheme, DeblurGAN-D, which uses a fine-tuned discriminator as a blur assessor to selectively deblur frames, and reports accuracy improvements for six of seven tested trackers.","tokens_in":14323,"tokens_out":2852,"duration_ms":30257,"significance":"If the empirical findings are valid, the paper makes a useful contribution: it provides a controlled benchmark for blur robustness, documents a non-monotonic effect of blur level on tracker accuracy, and proposes a lightweight, tracker-agnostic gating mechanism for deblurring. The breadth of the evaluation (23 trackers, 4 feature families, 500 videos) and the reproducibility-friendly design (built on the public NfS dataset) are strengths. The paper also honestly identifies limitations of prior benchmarks. However, the central quantitative conclusions rest on a specific ground-truth generation rule for blurred frames, and several evaluation protocols are oracle-based or lack statistical safeguards; these issues must be resolved before the benchmark's findings can be taken as established.","major_comments":[{"comment":"The ground-truth rule for blurred frames is not geometrically correct for moving targets. The blurred frame is defined as \\tilde I_t^L = avg(I_t,...,I_{t+L-1}) and the ground truth is 'the average of annotations of medium frames.' For a target that moves by d pixels during the exposure window, the visible smear occupies width w+d, while the averaged box has width approximately w and a centered midpoint. A tracker that perfectly predicts the true blurred target extent would receive IoU roughly w/(w+d), which decreases as L increases. Consequently, the AUC declines on S8/S16 and the conclusion 'heavy blur always hurts' may reflect the ground-truth rule rather than genuine tracker failure. The NRS metric inherits this bias because it gates frames using the IoU > 0.5 condition from S1 and then applies that gate to the undersized S{2,4,8,16} boxes. Please either justify the box-averaging rule with a physical model, use a union/coverage-based annotation for the integrated target, or validate the rule on human-annotated real blurred videos.","section":"3.1 (Dataset)"},{"comment":"The selective-deblurring results reported in Section 4.2 are oracle-based and cannot support the claim that selective deblurring 'improves the tracking performance of all trackers significantly.' The text states that 'the result with higher precision is saved as the final output,' where precision is computed from ground-truth center localization errors. This requires the ground truth at test time, so the gains in Figure 6 are an upper bound, not an achievable tracking algorithm. Please relabel this experiment as an oracle/upper-bound study and separate it from the achievable gains of the proposed DeblurGAN-D scheme in Section 5.3.","section":"4.2 (Pros of selective deblurring)"},{"comment":"No statistical significance tests or error bars are reported for any of the AUC comparisons. Claims such as '17 and 14 trackers have positive gains' on S2/S4 and 'light motion blur helps most of the compared trackers' are based on single benchmark runs; a paired bootstrap or Wilcoxon test across the 100/80 videos is needed to establish that these trends are not noise. The same applies to Section 5.4, where relative improvements (e.g., 9.3% for BT) are reported without variance or significance. Without such tests, the aggregate claims about 'most trackers' are not yet supported.","section":"4 and 5 (Evaluation methodology)"},{"comment":"There is a distributional leakage concern in the validation of the proposed scheme. The DeblurGAN-D assessor is fine-tuned on 20 scenes of the BVT benchmark (80 blurred videos generated by the same frame-averaging procedure) and then evaluated on the remaining 80 scenes of the same benchmark, generated by the same averaging process. This demonstrates improvement on the synthetic blur distribution used for training, but it does not establish that the assessor generalizes to real camera motion blur or to deblurring artifacts from other generators. Please add an evaluation on real blurred videos (e.g., the OTB motion-blur subset or GoPro-captured sequences) and report sensitivity to the threshold theta = 2.5.","section":"5.2 / 5.4 (Fine-tuning and evaluation split)"}],"minor_comments":[{"comment":"In the paragraph on AUC gains, the sentence 'Such numbers reduce to 7 and 2 on heavily-blurred subsets, i.e. S2 and S4' should refer to S8 and S16, not S2 and S4, which are the lightly blurred subsets.","section":"4.1"},{"comment":"The citation for BACF is incorrect: 'BACF [15]' should cite Galoogahi et al. 2017 (reference [16] in the bibliography), not the NfS benchmark [15].","section":"4.1"},{"comment":"The term 'full deburring' is used in several places; this appears to be a typo for 'full deblurring.' Please correct it for clarity.","section":"4.2 / 5"},{"comment":"The captions of Figures 8 and 9 mention 'bird sequences' and 'airplane sequences,' but the text does not identify which dataset or scenes these come from. Please clarify the source and the display convention (e.g., raw discriminator outputs vs. normalized scores).","section":"5.2"},{"comment":"Equation (4) defines P(I_t | s_t) as proportional to |D(hat I_t) - D(I_t)|, but this expression does not depend on s_t at all. Please specify how the two values of s_t (0 and 1) are represented in this likelihood, or revise the notation.","section":"5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant and timely question, and the proposed DeblurGAN-D scheme is interesting. However, the ground-truth averaging rule in Section 3.1 is a load-bearing methodological choice that may invalidate the central 'heavy blur always hurts' claim, and the Section 4.2 selective-deblurring gains are obtained with an oracle that uses ground truth. These issues are fixable in principle, but the manuscript needs substantial revision and re-evaluation before it can be accepted. I would also encourage the authors to release the benchmark annotations and evaluation code to support reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper delivers a genuinely new controlled benchmark for motion blur in tracking, built by averaging high-frame-rate NfS frames at five levels, and a practical trick: use DeblurGAN's discriminator score as a blur assessor to decide when to deblur an incoming frame. Both are worth engaging with. But the benchmark's ground-truth rule is the weak link, and I think the stress-test note is right: it may be taxing heavy-blur frames too harshly.\n\nWhat is new and good. The BVT benchmark is a sensible way to vary blur while keeping content fixed; evaluating 23 trackers on it gives a useful map of who degrades and who survives. The observation that light blur (L=2,4) often helps trackers is new and not just a rehash of existing benchmarks. The DeblurGAN-D assessor is a modest but genuine technique: instead of deblurring every frame, they use the difference in discriminator responses between raw and deblurred crops as a blur-level signal, then gate. On their held-out 80 scenes, it improves 6 of 7 trackers, which is respectable.\n\nWhere it gets soft. Section 3.1 sets the ground truth for a blurred frame as the average of the object boxes across the exposure window. When the target moves by d during those L frames, the visible smear has width w+d but the averaged box has width w. A tracker that predicts the smear perfectly scores IoU w/(w+d), which falls as L grows. So the AUC decline on S8 and S16 and the heavy-blur-always-hurts conclusion may be an artifact of the annotation rule rather than a real tracker failure. The NRS metric inherits this because it reuses the same boxes. The authors might intend the blurred frame to correspond to the object's instantaneous location at the middle of the exposure, which is a legitimate convention, but they never say that or defend it, and the OTB-style convention for blurred frames is the visible object box. This needs to be addressed head-on, e.g., by re-running key comparisons under a union-box GT or showing the conclusions are robust to that choice.\n\nEverything else is more routine. No error bars or significance tests on the AUC comparisons, and some gains are just a few percent. Section 4.2's selective deblurring is oracle-based (uses GT to pick the better box); the proposed method is separate and evaluated properly, but the 'significant' phrasing is borrowed from the oracle experiment. No code or data release, which is a bigger deal for a benchmark paper.\n\nBottom line: this is a solid, useful paper for tracking researchers, and it deserves peer review. It needs a revision that grapples with the GT-rule question and adds basic statistical rigor. I would not desk reject it.","headline":"Useful controlled blur benchmark and a smart deblurring gate, but the ground-truth rule for blurred frames may be biasing the central 'heavy blur hurts' finding.","tokens_in":14833,"tokens_out":7237,"would_cite":true,"duration_ms":71594,"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":"Motion blur is not uniformly bad for trackers: light blur aids many, heavy blur hurts, and a learned gate that selectively deblurs improves six existing trackers.","keywords":["motion blur","visual object tracking","benchmark","deblurring","GAN","blur robustness","selective deblurring","DeblurGAN"],"falsifier":"Run a suite of trackers on real blurred video recorded with a paired high-speed camera, so the true sharp target position is known at the blurred frames' exposure midpoint. If light real blur never raises AUC above the sharp baseline, or if the DeblurGAN-D gate does not select deblurred frames more often on heavy blur, the central claims are contradicted. A simpler check: replace the averaged-annotation ground truth with manually labeled blurred boxes on the BVT videos and see whether the benchmark conclusions change.","tokens_in":13913,"feed_emoji":"🎯","tokens_out":3640,"duration_ms":33205,"temperature":0.7,"pith_summary":"The paper argues that the effect of motion blur on visual object tracking is not monotonic: light blur often helps, heavy blur always hurts. To show this, it builds the Blurred Video Tracking (BVT) benchmark, 500 videos from 100 scenes at five blur levels produced by averaging successive 240 fps frames, with ground truth from averaged annotations. Twenty-three trackers are evaluated on this benchmark. The paper then shows that applying deblurring to every frame helps only on heavily blurred videos and hurts on lightly blurred ones. Its proposed scheme, DeblurGAN-D, fine-tunes the DeblurGAN discriminator to measure blur level and selectively deblur frames, which improves the accuracy and robustness of six trackers.","feed_headline":"A blur gate boosts six trackers; light blur can help tracking","feed_subtitle":"Five-level blur benchmark shows trackers are non-monotonic in blur, so deblurring should be selective, not automatic.","key_machinery":"The load-bearing objects are the benchmark construction and the blur assessor. Blurred videos are generated by averaging $L$ successive frames from 240 fps NfS videos ($L=1,2,4,8,16$), then temporally sampling every 8 frames to 30 fps; ground-truth boxes are the average of the annotations of the medium frames in each averaging window. The assessor is the DeblurGAN discriminator, fine-tuned on blur/deblur pairs from 20 benchmark scenes with the generator fixed, so it can rank blur levels. Its difference in scores between a deblurred and raw frame, $|D(\\hat{I}_t) - D(I_t)| > \\theta = 2.5$, triggers selective deblurring inside a Bayesian formulation with a selector variable.","core_discovery":"The central claim is that tracker robustness to motion blur is level-dependent and that deblurring should be gated by a learned blur assessor rather than applied blindly. On the BVT benchmark, 17 of 23 trackers gain AUC on the lightest blurred subset and 14 on the second-lightest, while nearly all lose accuracy on the heaviest. Full-frame deblurring with DeblurGAN or SRN lowers accuracy on light blur and raises it on heavy blur. The paper's method treats the fine-tuned discriminator's score difference $D(\\hat{I}_t) - D(I_t)$ as a blur-level signal; when it exceeds a threshold, the deblurred search region is used, otherwise the raw frame is kept. This selective scheme improves 6 of 7 tested trackers, with relative gains up to 9.3% for BT.","pith_inferences":["If light blur genuinely helps, a testable extension is to intentionally add small synthetic blur during tracker training or template augmentation; the paper's data suggests this could improve sharp-video accuracy for some trackers.","The same discriminator-difference gate could apply to other video tasks such as detection or segmentation, where heavy blur also corrupts features but light blur may not.","Temporal averaging of high-frame-rate frames may not capture all real camera blur, such as rolling-shutter or spatially varying kernels; validating the gate on real blurred video with paired sharp references would strengthen the transfer."],"forward_implications":["Tracker rankings on blur subsets are incomplete unless the blur level is controlled; a tracker can be best on sharp frames and worse than others under heavy blur.","Light blur acts as a mild regularizer or augmentation for many trackers, so blur robustness should be reported across levels rather than as a single score.","Full-frame deblurring is the wrong default; deblurring should be applied only when the assessor indicates heavy blur.","Existing trackers can be upgraded without retraining them, by wrapping them with the selective deblurring scheme.","The normalized robustness score (NRS) gives a blur-robustness measure that is separated from absolute accuracy."],"supporting_citations":[{"why":"Supplies the NfS dataset of 240 fps sharp videos and annotations used to generate the BVT blurred videos.","marker":"[15]"},{"why":"Provides DeblurGAN, whose generator deblurs frames and whose discriminator is fine-tuned into the blur assessor.","marker":"[28]"},{"why":"Provides SRN, the second deep deblurring method used to compare full and selective deblurring effects.","marker":"[51]"},{"why":"Defines the success plot and AUC metric that the BVT benchmark adapts for its three evaluation metrics.","marker":"[58]"},{"why":"Supplies ECO, a representative deep-feature tracker evaluated and improved by the selective deblurring scheme.","marker":"[6]"},{"why":"Supplies STRCF, a correlation-filter tracker evaluated and improved by the proposed scheme.","marker":"[31]"},{"why":"Supplies SiamFC, a Siamese tracker evaluated and improved by the proposed scheme.","marker":"[4]"},{"why":"Supplies MBT, a motion-blur-aware tracker used as a baseline that benefits less from selective deblurring.","marker":"[35]"}],"fun_headline_variants":["Light blur helps trackers; deblur only when heavy","Selective deblurring beats blanket deblurring for trackers","Tracking: light blur often helps; heavy blur always hurts","Gated deblurring boosts 6 trackers; skip it on light blur"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark's ground truth assumes that averaging the annotations of the sharp frames inside an exposure window gives the correct bounding box for the averaged blurred frame, and that averaging L successive sharp frames faithfully reproduces real motion blur; if either fails, the measured blur effects and the six-tracker improvement may not transfer to real blurred video.","fun_headline_variants_meta":{"raw":{"variants":["Light blur helps trackers; deblur only when heavy","Selective deblurring beats blanket deblurring for trackers","Tracking: light blur often helps; heavy blur always hurts","Gated deblurring boosts 6 trackers; skip it on light blur"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000909,"raw_usage":{"total_tokens":3911,"prompt_tokens":950,"completion_tokens":2961,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":2883}},"tokens_in":566,"tokens_out":2961,"duration_ms":18793,"temperature":1.0,"reasoning_tokens":2883,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:53:45.935289+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a suite of trackers on real blurred video recorded with a paired high-speed camera, so the true sharp target position is known at the blurred frames' exposure midpoint. If light real blur never raises AUC above the sharp baseline, or if the DeblurGAN-D gate does not select deblurred frames more often on heavy blur, the central claims are contradicted. A simpler check: replace the averaged-annotation ground truth with manually labeled blurred boxes on the BVT videos and see whether the benchmark conclusions change.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NfS dataset of 240 fps sharp videos and annotations used to generate the BVT blurred videos."},{"cited_title":"Kupyn, V","cited_arxiv_id":null,"evidence_quote":"Provides DeblurGAN, whose generator deblurs frames and whose discriminator is fine-tuned into the blur assessor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides SRN, the second deep deblurring method used to compare full and selective deblurring effects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the success plot and AUC metric that the BVT benchmark adapts for its three evaluation metrics."},{"cited_title":"Danelljan, G","cited_arxiv_id":null,"evidence_quote":"Supplies ECO, a representative deep-feature tracker evaluated and improved by the selective deblurring scheme."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies STRCF, a correlation-filter tracker evaluated and improved by the proposed scheme."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies MBT, a motion-blur-aware tracker used as a baseline that benefits less from selective deblurring."}],"review_version":1}