{"id":"768bfb88-515f-441e-b30d-9f589dff97e2","arxiv_id":"1909.00319","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A long-term tracker combining MDNet and SiamRPN++ with a visibility judgement module and a cascade re-detection module reports higher VOT and VisDrone scores than its short-term baselines.","lead":"This paper describes a long-term object tracking algorithm that couples two short-term trackers with a module that decides whether the target is still visible, plus a cascade detector that searches for the target after it reappears. The idea matters for drones, surveillance, and robots, where targets routinely leave the field of view and must be re-acquired.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The hand-set judgement thresholds in Section 3.3 are scene-dependent by the authors' own admission, and no sensitivity analysis or ablation isolates their contribution, so the reported LTB35 and VisDrone gains may not generalize.","rationale":"The paper is an engineering integration with plausible components and internally consistent reported numbers; there is no formal circularity or obvious contradiction. The load-bearing weak point is the unvalidated operating point in the decision module, which the authors themselves flag as scene-dependent. A threshold sweep would settle it. Since the reader's verdict is already CONDITIONAL, with a correct weakest_assumption, the appropriate stress-test outcome is to keep that verdict. I do not see evidence to raise or lower it further; the absence of code and error bars are secondary but do not by themselves undermine the argument enough to move to REJECT.","tokens_in":836,"tokens_out":1735,"duration_ms":73213,"concrete_test":"Re-run the VOT-2018 LTB35 evaluation with all components fixed except theta_mid swept over {0.3, 0.4, 0.5, 0.6} and theta_low swept over {0.05, 0.1, 0.2}, producing an F-score surface. If the reported F-score changes by more than roughly 0.05 across these neighboring settings, the result is dominated by hand-set thresholds; if the surface is flat, the concern is not strongly supported. Also run an ablation with the detection module disabled to confirm the gain over MDNet is due to re-detection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that combining MDNet, SiamRPN++ and a cascade detector with an optical-flow-guided judgement improves long-term tracking. This depends on Section 3.3's hand-set thresholds theta_mid = 0.5 and theta_low = 0.1, which decide whether to accept a track, correct it, or trigger re-detection. The authors themselves, at the end of the VisDrone validation paragraph in Section 4.2, note that a pre-defined failure-judgement threshold is 'closely related to the scene' when explaining MBMD's weakness; the same applies to their own thresholds. No calibration, sensitivity analysis, or ablation is reported. If theta_mid and theta_low are not appropriate for a given sequence, the cascade detector either starts too early (false positives) or too late (missed reappearance), so the F-score gains on LTB35 and the VisDrone success/precision gains could be an artifact of the operating point rather than of the proposed flow-guided cascade. The paper also omits a comparison against a strong long-term baseline such as MBMD in Table 1, weakening the 'state-of-the-art' claim. The decisive missing experiment is a threshold sweep and an ablation that turns off the judgement/detection modules.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes flow_MDNet_RPN, a long-term single-object tracker that couples two short-term trackers (MDNet and SiamRPN++) with a rule-based judgement module and a cascade detection module. The judgement module uses MDNet classification scores, SiamRPN++ similarity scores, and PWC-Net optical flow to decide whether the target is present, and the detection module expands the search area from local to global when the target is deemed absent. The method is evaluated on VOT-2018 LTB35 and VisDrone-SOT2019 validation and test sets, reporting an F-score of 0.5405 on LTB35 and success/precision improvements over MDNet and SiamRPN++ on VisDrone.","tokens_in":10756,"tokens_out":3900,"duration_ms":31805,"significance":"If the reported results are reproducible, the paper provides a useful engineering recipe for converting short-term trackers into long-term ones: the combination of two complementary short-term trackers with a flow-guided re-detection mechanism gives consistent gains over its two baselines on long-term benchmarks. The method is described in enough detail for reimplementation, and the evaluation uses public benchmarks with standard metrics. However, the state-of-the-art claim is not backed by comparison with strong long-term baselines in the LTB35 table, and the central decision thresholds are not validated; the significance is therefore conditional on additional evidence.","major_comments":[{"comment":"The judgement thresholds theta_mid=0.5 and theta_low=0.1 are hand-set, and no sensitivity analysis or ablation is reported. The authors themselves note in Section 4.2 that pre-defined failure-judgement thresholds are 'closely related to the scene' in their discussion of MBMD, which applies equally to their own thresholds. Because these thresholds control whether the cascade detector is triggered, the reported F-score and success/precision gains could be an artifact of the chosen operating point rather than of the flow-guided cascade. A threshold sweep and an ablation that turns off the judgement/detection modules are needed to support the central claim.","section":"3.3, 4.2"},{"comment":"Table 1 compares only SiamRPN++ and MDNet on VOT-2018 LTB35, yet Section 4.2 states that the proposed tracker 'achieves state-of-the-art performance' on this dataset. Without comparisons to other published long-term trackers on LTB35 (e.g., MBMD, DaSiamRPN_LT, TLD, LCT), the state-of-the-art claim is unsupported. The comparison with MBMD in Figure 5 is limited to VisDrone validation and is only described qualitatively in the text; numerical results for that figure or a supplementary table would strengthen the claim.","section":"4.2, Table 1"},{"comment":"The sentence 'our proposed algorithm which has been trained on the VisDrone-SOT2019 training dataset is more suitable for video taken by drones' suggests that the method or its components were trained on VisDrone training data before evaluation on the validation and test splits. This is not described in the implementation details (Section 4.1), which list only ImageNet/COCO/YouTube-BB pretraining. If VisDrone training data was used, it should be disclosed explicitly, since it affects the fairness of comparisons against trackers not trained on VisDrone; if not, the sentence should be corrected.","section":"4.2"},{"comment":"The reported numbers are single-run results despite the stochastic nature of MDNet's Gaussian sampling and online updates; no error bars, multiple-run statistics, or variance estimates are provided. Since several reported improvements (e.g., 5.2 success points over MDNet on VisDrone validation) are modest relative to likely run-to-run variance, the quantitative claims need at least a statement of variance or a deterministic protocol.","section":"4.2, Table 1"}],"minor_comments":[{"comment":"The sentence 'Finally, we summarize our work in Section 1' should refer to Section 5 (Conclusion).","section":"1"},{"comment":"Equation (1) is garbled by duplicated lines and likely a rendering issue; please fix the formula and define the threshold theta_F explicitly.","section":"4.2, Eq. (1)"},{"comment":"The flowcharts and framework figures are low-resolution, and the symbols for the scores are barely legible; please provide higher-quality versions.","section":"Figures 2-4"},{"comment":"The notation switches between 'classification score s_c' and 'similarity score s_s' and later uses 'confidence s_t' without a definition; please define all scores and their ranges (e.g., whether MDNet scores are probabilities or raw classifier outputs).","section":"3.3"},{"comment":"Several references have incomplete author lists (e.g., [9]) and inconsistent formatting; please check the bibliography.","section":"References"},{"comment":"In the VisDrone validation paragraph, the text says 'ours is much higher than the performance of the baseline algorithm MDNet and SiamRPN++' but does not provide numerical results for TLD, LCT, and MBMD in a table; adding the actual values would help the reader.","section":"4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a modest but plausible engineering contribution. The main risks are the unvalidated thresholds and the incomplete comparison set, both of which are addressable in revision. The use of 'state-of-the-art' should be toned down unless the comparison is expanded. No issues of attribution or novelty disclosure beyond the normal expectations for a workshop-level contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering paper that combines MDNet, SiamRPN++, PWC-Net, and GA-RPN into a long-term tracker. The design pattern—short-term tracker plus failure judgement plus re-detection—is familiar from MBMD, TLD, and DaSiamRPN_LT, so the novelty is in the specific fusion and the optical-flow-guided cascade, not in a new mechanism. The reported numbers are internally consistent, and the authors are candid about the weak spots in their own method. But the evidence does not support the 'state-of-the-art' statement as written, and the missing experiments keep me from trusting the gains without more work.\n\nWhat is new: the concrete combination of MDNet's online updating with SiamRPN++'s stable similarity, a score-based judgement rule with two thresholds, and a three-stage local-to-global cascade using flow compensation and GA-RPN proposals. The method is specified clearly enough to reproduce, and the figures help.\n\nWhat it does well: it evaluates on VOT-LTB35 and VisDrone, and the VisDrone comparison includes a reasonable set of baselines. The authors note that their precision is lower than SiamRPN++ because detection errors bleed into tracking, which is a fair self-assessment. They also concede that the fixed failure threshold is scene-dependent—the same criticism they level at MBMD.\n\nSoft spots, in order of severity. First, the LTB35 table contains only SiamRPN++ and MDNet; no MBMD, TLD, or LCT. The state-of-the-art claim rests on that two-row comparison. Second, there are no error bars, multiple runs, code, or an ablation that isolates the contribution of the judgement and detection modules. The thresholds theta_mid = 0.5 and theta_low = 0.1 are hand-chosen, and the paper's own admission that thresholds are scene-dependent makes the lack of sensitivity analysis a real gap. Third, the authors state in the VisDrone validation section that their algorithm 'has been trained on the VisDrone-SOT2019 training dataset.' If that means they trained or tuned on the same benchmark's training split while comparing against published baselines that did not, the comparison is not apples-to-apples. That needs clarification. Fourth, the writing is rough: the F-measure equation is garbled and the text has many OCR-style artifacts. That is not a scientific flaw, but it adds friction.\n\nThe citation pattern looks appropriate: they cite the modules they use and the benchmarks they evaluate on. No obvious self-citation problem.\n\nBottom line: the central argument—add a judgement module and a cascade detector to a short-term tracker—is sound and the reported gains are plausible. This is a paper for practitioners building long-term trackers for drones and surveillance, who will find the integration recipe useful. It deserves a serious referee, not a desk reject, but it needs major revision before the claims can be accepted. I would send it to review and ask specifically for the baseline additions, a threshold sweep, an ablation, and clarification of the VisDrone training data use.","headline":"Competent integration of MDNet, SiamRPN++, and a flow-guided cascade, but the state-of-the-art claim outruns the evidence.","tokens_in":11281,"tokens_out":4579,"would_cite":false,"duration_ms":39224,"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":"A long-term tracker built from MDNet, SiamRPN++, and a flow-guided cascade detector can re-catch targets after occlusion or out-of-view events, reporting 0.5405 F-score on VOT-2018 LTB35 and large gains over its MDNet baseline on…","keywords":["long-term object tracking","single object tracking","MDNet","SiamRPN++","cascade detection","optical flow","VOT-2018 LTB35","VisDrone-SOT2019"],"falsifier":"Disable the cascade detection module while keeping the same judgement logic, then rerun on VisDrone-SOT2019 test; if the 10.4 success-point and 17 precision-point gains over MDNet persist, re-detection is not what carries the result. Alternatively, sweep the two fixed thresholds per sequence; if no single setting reproduces the reported F-score on LTB35, the fixed rule is not the mechanism.","tokens_in":10270,"feed_emoji":"🎯","tokens_out":6410,"duration_ms":53895,"temperature":0.7,"pith_summary":"Long-term visual tracking requires knowing not only where the target is, but also when it has disappeared and how to catch it when it reappears. This paper claims that a standard short-term tracker can be turned into an effective long-term tracker by adding a judgement module and a cascade detection module. The judgement module decides, from two complementary scores, whether the reported box actually contains the target; the cascade detection module expands its search from a local neighborhood to the whole frame, guided by optical flow, and re-locates the target after occlusion or out-of-view events. On the VOT-2018 LTB35 benchmark the combined system reaches an F-score of 0.5405, and on VisDrone-SOT2019 it improves over the MDNet baseline by 10.4 success points and 17 precision points on the test set.","feed_headline":"Cascade detector lets long-term tracker recapture lost targets","feed_subtitle":"Reported gains: 0.5405 F-score on VOT-2018 LTB35 and 10.4 success points over MDNet on VisDrone test.","key_machinery":"The load-bearing object is the three-module pipeline: a short-term tracking module (MDNet plus SiamRPN++), a judgement module, and a cascade detection module. The judgement module works on two numbers: the MDNet classification score $S_c$ and the SiamRPN++ similarity score $S_s$, compared against thresholds $\\theta_{\\mathrm{mid}}=0.5$ and $\\theta_{\\mathrm{low}}=0.1$. Depending on which score is above or below threshold, the algorithm either accepts the box, resamples around it with Gaussian sampling, compensates camera motion with optical flow from PWC-Net, or regresses the box. When both scores fall below the low threshold, the cascade detector starts: it searches locally with flow-guided sampling, then with GA-RPN proposals in a $5^2$-area patch, then an $18^2$-area patch, then the whole frame, until both scores clear the threshold. This mechanism is what lets the tracker survive long out-of-view intervals.","core_discovery":"The central claim is that target disappearance can be handled by close cooperation between a short-term tracker, a failure judge, and a detector, rather than by building a one-stage long-term model. The paper assembles a short-term module from two complementary trackers: MDNet contributes an online-updated classifier with a classification score, and SiamRPN++ contributes a fixed Siamese similarity score. A rule-based judgement module combines these scores and, on failure, hands control to a cascade detector that first searches around the last position with Gaussian sampling and optical-flow compensation, then widens to a proposed-box search and finally to the global frame. The reported result is that this pipeline, called flow_MDNet_RPN, achieves the best F-score on VOT-2018 LTB35 among the compared trackers and large success and precision gains over its MDNet baseline on VisDrone-SOT2019 test sequences.","pith_inferences":["Extension: an untested improvement implied by the design is learning the judgement thresholds per scene or adapting them online; the paper itself notes the fixed thresholds are closely related to the scene.","Extension: because the detector is activated only by the fixed thresholds, the same pipeline could be combined with a learned target-presence classifier to remove the manual threshold choice.","Extension: on sequences with frequent distractors, the SiamRPN++ similarity score may need a higher threshold to avoid false re-detection, so the reported 0.6095 precision on LTB35 could degrade on denser scenes."],"forward_implications":["The proposed flow_MDNet_RPN reports both a target box and a presence confidence, so it is suited to practical tasks that need to know when the target is absent.","The cascade detection module expands from local to global search, so the tracker can recover targets that reappear far from their last location.","Optical-flow compensation in the judgement and detection steps is intended to handle camera motion, a common failure source in drone videos.","On the reported benchmarks, the largest gaps over MDNet and SiamRPN++ come from sequences where the target leaves view, suggesting re-detection is the main source of improvement."],"supporting_citations":[{"why":"Supplies the MDNet short-term tracker, the classification score $S_c$, and the baseline whose VisDrone numbers the paper improves.","marker":"[4]"},{"why":"Supplies SiamRPN++, the Siamese similarity score $S_s$, and the complementary short-term tracker that jointly localizes candidates.","marker":"[22]"},{"why":"Supplies PWC-Net, the optical flow network used to compensate global motion in both the judgement and cascade detection stages.","marker":"[32]"},{"why":"Supplies GA-RPN, the region proposal network that generates candidate boxes for the expanded local and global search areas.","marker":"[33]"},{"why":"Provides the VOT2018-LT benchmark and the F-score, precision, and recall protocol on which the main 0.5405 result is reported.","marker":"[9]"},{"why":"Provides the VisDrone-SOT2019 dataset and the success and precision evaluation protocol used for validation, testing, and attribute analysis.","marker":"[10]"},{"why":"MBMD is the long-term tracker with re-detection that the paper compares against and claims to surpass on VisDrone.","marker":"[27]"}],"fun_headline_variants":["Long-term tracker recaptures lost targets via cascade detection","Cascade detection lets long-term tracker reacquire vanished targets","Flow-guided trackers plus cascade detection fix long-term tracking","When targets vanish, cascade detection keeps long-term tracker on track"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole gain rests on two fixed numbers in the judgement module, 0.5 and 0.1, reliably telling apart frames where the target is visible from frames where it is gone; if they misfire, the detector starts too early or too late and the improvement disappears.","fun_headline_variants_meta":{"raw":{"variants":["Long-term tracker recaptures lost targets via cascade detection","Cascade detection lets long-term tracker reacquire vanished targets","Flow-guided trackers plus cascade detection fix long-term tracking","When targets vanish, cascade detection keeps long-term tracker on track"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":2996,"prompt_tokens":839,"completion_tokens":2157,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":2088}},"tokens_in":455,"tokens_out":2157,"duration_ms":16175,"temperature":1.0,"reasoning_tokens":2088,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:55:06.772674+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Disable the cascade detection module while keeping the same judgement logic, then rerun on VisDrone-SOT2019 test; if the 10.4 success-point and 17 precision-point gains over MDNet persist, re-detection is not what carries the result. Alternatively, sweep the two fixed thresholds per sequence; if no single setting reproduces the reported F-score on LTB35, the fixed rule is not the mechanism.","supporting_citations":[{"cited_title":"Nam, and B","cited_arxiv_id":null,"evidence_quote":"Supplies the MDNet short-term tracker, the classification score $S_c$, and the baseline whose VisDrone numbers the paper improves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies SiamRPN++, the Siamese similarity score $S_s$, and the complementary short-term tracker that jointly localizes candidates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies PWC-Net, the optical flow network used to compensate global motion in both the judgement and cascade detection stages."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GA-RPN, the region proposal network that generates candidate boxes for the expanded local and global search areas."},{"cited_title":"Kristan, A","cited_arxiv_id":null,"evidence_quote":"Provides the VOT2018-LT benchmark and the F-score, precision, and recall protocol on which the main 0.5405 result is reported."}],"review_version":1}