{"id":"f7e76fdd-c9a1-439f-a6f0-9b08913eec11","arxiv_id":"2411.10028","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MOT FCG++ adds diagonal-modulated GIoU, dynamic appearance, and averaged constant velocity modeling to MOT FCG, improving HOTA, MOTA, and IDF1 on MOT17 and MOT20.","lead":"This paper presents MOT FCG++, a multi-object tracker that improves the clustering-based MOT FCG with three tweaks: a diagonal-modulated overlap score, a confidence-weighted appearance memory, and an averaged velocity model. The result is a few points better on MOT17 and MOT20 benchmarks, with the largest gains in identity metrics like IDF1.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The baseline in Table 1 is not a controlled comparator: Section 4.1 fine-tunes the YOLOX detector and swaps in SBS ReID features, so the reported module gains are conflated with detector/embedding changes until a counterfactual re-run of MOT FCG is shown.","rationale":"Read in good faith: the paper's stated goal is an incremental engineering improvement over MOT FCG. The modules are technically plausible; dynamic appearance is a reasonable fix for median representation, and averaging velocity over N frames is consistent with reducing observation noise. The ablation directions are mostly monotone and the selection studies are useful. I do not think the paper is fraudulent or that the central claim is impossible. However, the most defensible version of the central claim is that these three modules improve MOT FCG by the reported amounts. That claim requires a controlled comparison in which the baseline uses exactly the same detector and ReID features. The paper explicitly changes two components before the comparison and never reports a re-run of MOT FCG with those components. The ablation baseline reproduces the published MOT FCG values exactly, so either the tracker is deterministic and the detector/ReID setup is identical, or the baseline row is reused; in either case the paper does not provide an independent re-run under the FCG++ configuration. The test-set deltas are small enough that detector fine-tuning alone could account for them. This is not a disagreement with external consensus; it is an internal control problem. The right remedy is conditional acceptance pending a controlled re-run or release of code enabling one. I therefore agree with the reader's conditional verdict and do not propose a stronger one.","tokens_in":11596,"tokens_out":6418,"duration_ms":64720,"concrete_test":"Run a controlled counterfactual in the same codebase: use the exact fine-tuned YOLOX detections and SBS features adopted for FCG++, then disable only DGIoU (revert to IoU), dynamic appearance (revert to median), and ACV (revert to two-frame constant velocity). Report HOTA/MOTA/IDF1 on MOT17-val, MOT20-val, and DanceTrack-val for (a) this no-module baseline, (b) FCG++, and (c) optionally MOT FCG with the un-fine-tuned detector. If the no-module baseline matches the published MOT FCG numbers within noise, the Table 1 deltas can be attributed to the modules; if it drifts toward the FCG++ numbers, the headline comparison overstates module gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the three proposed modules (DGIoU, dynamic appearance, ACV) generate the Table 1 gains over MOT FCG. The load-bearing assumption is that the only difference between the two columns is those modules. Section 4.1 breaks that assumption: \"we fine-tune the YOLOX detector used in MOT FCG, and we extract ReID features using the SBS network.\" The MOT FCG column is taken from the published paper, not re-run under FCG++'s detection/ReID configuration. Table 4's validation baseline also reproduces the published values (e.g., MOT17 HOTA 74.73/IDF1 80.54, MOT20 HOTA 65.71/IDF1 72.18), which suggests the ablation baseline was not recomputed with the same fine-tuned detector and SBS embeddings. Since detection quality and embedding quality directly enter the tracklet clustering, any improvement from the detector/ReID change is conflated with the module contributions. The reported deltas are small (+0.5 HOTA on MOT17, +2.6 IDF1 on MOT20); even modest detection differences can produce such shifts, so the attribution is not established by the data shown.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three modifications to the hierarchical clustering tracker MOT FCG: Diagonal Modulated GIoU (DGIoU) for spatial motion similarity, a dynamic confidence-weighted appearance representation, and Average Constant Velocity Modeling (ACV) for motion prediction. The authors report consistent improvements over MOT FCG on MOT17, MOT20, and DanceTrack, with headline numbers of 63.1 HOTA, 76.9 MOTA, and 78.2 IDF1 on the MOT17 test set. The paper includes benchmark comparisons and ablation studies on validation splits.","tokens_in":11863,"tokens_out":4127,"duration_ms":42096,"significance":"If the attribution were established, the work would offer a simple, incremental improvement to clustering-based MOT, with three small modules that are easy to understand and reproduce. The paper is honest about the modest size of the gains and validates on three benchmarks with complementary characteristics. The main strengths are the breadth of experiments and the clear separation of the three proposed components. However, the central claim that the proposed modules cause the reported gains is currently undermined by an uncontrolled baseline comparison and by a direct inconsistency between the ablation table and the main results table for MOT20.","major_comments":[{"comment":"The baseline is not a controlled comparator. Section 4.1 states that the YOLOX detector is fine-tuned and that ReID features are extracted with the SBS network, while the MOT FCG numbers are taken from the published paper. Since detection quality and embedding quality directly affect tracklet clustering, the reported deltas (+0.5 HOTA on MOT17, +2.6 IDF1 on MOT20) cannot be attributed to DGIoU, Dynamic Appearance, and ACV. The authors should re-run MOT FCG under exactly the FCG++ detector/ReID configuration and use that as the ablation baseline, or explicitly quantify the individual contributions of the detector/ReID changes.","section":"4.1, Tables 1 and 4"},{"comment":"The normalization statement for DGIoU is mathematically incorrect. The text claims that dividing dDGIoU by 2 ensures dGIoU is in [0,1], but dDGIoU = 1 - (L2/L1)*GIoU is not guaranteed to be in [0,1] because GIoU ranges over [-1,1] and L2/L1 is an arbitrary positive ratio. If lambda_C is intended as a bounded similarity or distance, its range should be defined carefully or clipped explicitly. This is a load-bearing detail for the DGIoU contribution.","section":"3.2, Eq. (1)"},{"comment":"There is a direct numerical inconsistency for MOT20. The final ablation row in Table 4 reports MOT20-val HOTA of 68.81, while Table 3 reports MOT FCG++ on the same validation split at 65.8 HOTA, and the text says the improvement is from 65.7 to 65.8. A three-point HOTA discrepancy is far larger than the claimed improvements and must be reconciled; as printed, the two tables cannot both describe the same configuration.","section":"Table 4 vs. Table 3"},{"comment":"The hyperparameter and ablation analysis is not fully convincing. In Section 4.3, N is selected as 9 based on HOTA on MOT17-train, but the maxima in Figure 7 differ across metrics (MOTA peaks at N=16, AssA at N=8), and no error bars or multiple-seed variance are reported. Additionally, beta_f and off are tuned per dataset on the same validation splits where the final numbers are reported, so the risk of overfitting to the benchmark is not quantified. Please report sensitivity to these choices and variance where feasible.","section":"4.3, Figure 7, Section 4.1"}],"minor_comments":[{"comment":"The caption reports '68.1 MOTA, 72.3 IDF1 and 58.4 HOTA on the MOT17-test set', but the numbers correspond to MOT20; the dataset name is a typo.","section":"Figure 1 caption"},{"comment":"The two subcaptions appear garbled ('with or Dynamic Appearance Feature', 'with or Median Appearance Feature') and do not clearly distinguish the panels; 'with/without' is presumably intended.","section":"Figure 6 captions"},{"comment":"The notation for the EMA weight is inconsistent: Eq. (2) uses beta as a constant, while Eq. (3) defines beta_t as adaptive; please clarify the update rule and the dependence of beta_t on the detection confidence.","section":"3.3, Eqs. (2)-(3)"},{"comment":"The parameter is called 'n' in the method text ('if the trajectory length is less than n') but 'N' in the ablation and equation; please use one symbol consistently.","section":"3.4 and 4.3"},{"comment":"The hyperparameter names are typeset inconsistently ('beta_f and of f' instead of 'beta_f and off'); also, 'off' is not defined beyond being an offset in Eq. (1), so its role and allowed range should be stated.","section":"Section 4.1"},{"comment":"Reference [32] is malformed ('et al Luiten, JonathonOsep'); please fix the author list and formatting.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental benchmark improvement and does not provide code or release of the trained models. The most serious issue is the lack of a controlled baseline: without a re-run of MOT FCG under the same detector and ReID configuration, the module attribution is not established. The MOT20 HOTA discrepancy between Table 3 and Table 4 also suggests a possible data-handling or labeling error that the authors must resolve. If the authors can provide the counterfactual baseline and reconcile the tables, the paper could become publishable after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest, engineering-oriented MOT paper. It takes three existing ideas—confidence-weighted EMA appearance (from Deep-OC-SORT), diagonal-modulated GIoU (from Hybrid-SORT's weak-cue line), and a windowed average velocity—and applies them to the clustering tracker MOT FCG. The motivation is clear, and the ablations on MOT17, MOT20, and DanceTrack mostly show small, consistent gains. But the paper never cleanly separates the proposed modules from a detector/ReID change, and several numbers disagree with themselves.\n\nThe main soft spot: Section 4.1 says the authors fine-tune the YOLOX detector and use SBS ReID features, while the baseline MOT FCG numbers are taken from the published paper. The ablation baseline in Table 4 reproduces the published MOT FCG values, so the module increments are measured against the original detector/ReID, not against the new components. The headline test-set deltas in Table 1 (+0.5 HOTA on MOT17, +2.6 IDF1 on MOT20) compare the full system to the original tracker, so those deltas conflate module gains with detector/embedding gains. Given how small the deltas are, that matters. The fix is simple: re-run MOT FCG with the same fine-tuned detector and SBS features as a proper counterfactual.\n\nAlso, Table 4 lists MOT20-val HOTA as 68.81 for the full model, while Table 3 and the text say 65.8. That is a three-point gap on the same dataset and cannot be brushed aside. The DGIoU normalization statement is mathematically wrong: dDGIoU/2 is not guaranteed to lie in [0,1]. The Figure 1 caption repeats 'MOT17-test' for both datasets. No error bars, no code, and beta_f/off are tuned per dataset on the same benchmarks where results are reported.\n\nWhat is genuinely useful: the diagnosis that MOT FCG's median-element appearance feature is fragile is correct, and the dynamic appearance fix is sensible. The DGIoU idea is a reasonable way to inject shape information into clustering association. The ablations, if they were run on the original detector/ReID, do suggest DA and DGIoU each help. So the core idea has some support; it is just not packaged cleanly.\n\nWho is this for? Someone working on clustering-based or appearance-based MOT who wants a quick read on weak-cue injection. It is not a breakthrough, but it is not junk either.\n\nIf you are asked to referee, engage with it—the central question is empirical and could be settled with a proper counterfactual and corrected tables. I would not desk-reject it, but I would put the burden on the authors to release code and fix the inconsistencies.","headline":"Incremental MOT paper with a plausible weak-cue idea, but an uncontrolled baseline and inconsistent MOT20 numbers muddy the attribution.","tokens_in":12443,"tokens_out":6143,"would_cite":false,"duration_ms":54184,"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":"This paper claims that three small changes to how a tracker represents motion and appearance — a diagonally modulated GIoU, a confidence-weighted dynamic appearance embedding, and an averaged constant-velocity model — each improve the…","keywords":["multi-object tracking","hierarchical clustering","appearance embedding","motion features","GIoU","exponential moving average","MOT17","MOT20"],"falsifier":"Reproduce MOT FCG++ with the same fine-tuned YOLOX detector and SBS ReID features but swap DGIoU back to plain IoU; if HOTA on MOT17-val does not drop below the reported 76.14, the paper's attribution of the gain to the diagonal modulation fails.","tokens_in":1726,"feed_emoji":"🎯","tokens_out":4267,"duration_ms":53701,"temperature":0.7,"pith_summary":"This paper claims that a clustering-based multi-object tracker can be improved without new detectors or new network architectures, by refining how trajectory-level motion and appearance features are represented. Starting from MOT FCG, the authors replace plain IoU with a Diagonal Modulated GIoU that encodes both position and box shape, replace the median-element appearance feature with a confidence-weighted exponential moving average over the trajectory, and replace two-frame velocity estimates with a velocity averaged over the last N frames. They report consistent gains across MOT17, MOT20, and DanceTrack, with the largest reported improvement being a +2.6 IDF1 gain on the MOT20 test set. A sympathetic reader would care because the claim is that weak cues that existing clustering trackers discard — box diagonals, detection confidence, and multi-frame motion history — are enough to make association more robust in crowded and occluded scenes.","feed_headline":"Weak cues lift a tracking baseline to 63.1 HOTA on MOT17","feed_subtitle":"Three small changes to motion and appearance features beat MOT FCG on all reported metrics.","key_machinery":"The central machinery is the trio of representation upgrades layered onto MOT FCG's hierarchical clustering: (1) DGIoU, a spatial similarity score that multiplies GIoU by the ratio of the enclosing rectangle's side lengths, making the metric sensitive to box shape and relative scale; (2) Dynamic Appearance, an exponential moving average of ReID embeddings whose update weight decreases as detection confidence drops, so low-quality frames contribute less; and (3) ACV, a constant-velocity motion model that averages displacement over the last $N$ frames to suppress observation noise. The paper uses these as weak cues within the existing clustering framework, claiming that they make the trajectory-level features more global and more robust without changing the hierarchical association structure itself.","core_discovery":"The paper's central claim is that MOT FCG's two feature representations are the bottleneck: IoU misrepresents positional and shape relationships between bounding boxes, and the median-element appearance feature ignores both global trajectory information and detection quality. The proposed replacements are Diagonal Modulated GIoU, defined as $d_{\\mathrm{DGIoU}} = 1 - (L_2/L_1)\\cdot\\mathrm{GIoU}$, where $L_1,L_2$ are the box-side lengths entering the enclosing rectangle, so that the metric reflects aspect-ratio and scale; Dynamic Appearance, an EMA with adaptive weight $\\beta_t$ driven by detection confidence $s_{\\mathrm{det}}$; and Average Constant Velocity Modeling, $v = (x_t - x_{t-N})/N$, which averages displacement over the last $N$ frames instead of using the last two. The paper argues each module contributes a small but consistent gain, with DGIoU helping across all three datasets and Dynamic Appearance offering the largest identity-maintenance gains during occlusion.","pith_inferences":["A natural extension, not tested in the paper, is whether diagonal modulation generalizes to other spatial metrics such as center-distance or scaled IoU, since the success of $L_2/L_1$ suggests any cue carrying aspect-ratio information could help clustering association.","The reported optimum of N = 9 on MOT17 may be dataset-dependent; on DanceTrack, where motion is nonlinear and fast-changing, shorter averaging windows might be safer, and this is a directly testable hyperparameter sweep.","Because the paper's baseline numbers come from the published MOT FCG paper while its own runs use a fine-tuned YOLOX detector and SBS ReID features, the cleanest test of the three modules is to toggle each on and off inside a single codebase with identical detector and ReID weights; the paper's ablation does not fully isolate the modules from the component upgrades.","The confidence-threshold formulation of Dynamic Appearance suggests a natural robustness test: degrade detection confidence artificially and check whether the adaptive weight prevents identity switches better than a fixed-weight EMA."],"forward_implications":["If the claim holds, clustering-based trackers can be improved inexpensively by enriching the spatial metric with shape information, since DGIoU is reported to beat plain IoU and width- or height-modulated variants on MOT17.","The confidence-weighted dynamic appearance representation is a drop-in replacement for median-element or mean appearance features, and the paper reports it prevents identity switches during occlusion in a qualitative MOT17 example.","Averaging velocity over N frames stabilizes motion prediction; the paper reports N = 9 as the best setting on MOT17 and shows that the gain persists over a wide range of N.","The largest relative gain is on MOT20's dense, heavily occluded scenes, where IDF1 rises from 69.7 to 72.3, suggesting the improvements are especially relevant when appearance evidence is noisy.","With the optional StrongSORT post-processing (AFLink+GSI), the paper reports further gains to 64.1 HOTA and 79.2 MOTA on MOT17, indicating the modules compose with downstream trajectory smoothing."],"supporting_citations":[{"why":"MOT FCG is the baseline hierarchical clustering tracker whose median-element appearance and IoU motion features this paper replaces.","marker":"[24]"},{"why":"Hybrid-SORT supplies the weak-cue idea and the L2/L1 diagonal modulation term used to define DGIoU.","marker":"[25]"},{"why":"Deep-OC-SORT supplies the Dynamic Appearance representation with confidence-adaptive EMA weighting that the paper adapts for clustering tracklets.","marker":"[23]"},{"why":"YOLOX is the detector the paper fine-tunes for its private-detection protocol.","marker":"[34]"},{"why":"SBS (FastReID) is the network used to extract the ReID appearance embeddings.","marker":"[36]"},{"why":"MOT17 is the main benchmark dataset on which the headline HOTA, MOTA, and IDF1 numbers are reported.","marker":"[27]"},{"why":"StrongSORT's AFLink+GSI post-processing is applied to produce the MOT FCG++* results, demonstrating compatibility with trajectory smoothing.","marker":"[5]"}],"fun_headline_variants":["MOT FCG++: better motion and appearance cues reach 63.1 HOTA","Three tweaks to motion and appearance features lift MOT FCG to 63.1 HOTA","Aspect-aware GIoU and dynamic appearance lift MOT FCG to 63.1 HOTA","Improving on MOT FCG: new features yield 63.1 HOTA on MOT17","Diagonal GIoU and dynamic EMA sharpen multi-object tracking to 63.1 HOTA"],"cache_read_input_tokens":14464,"weakest_assumption_plain":"The claim that the three modules cause the improvements depends on the baseline MOT FCG numbers being directly comparable to the new runs, but the new runs use a fine-tuned detector and a different ReID network than the published baseline did.","fun_headline_variants_meta":{"raw":{"variants":["MOT FCG++: better motion and appearance cues reach 63.1 HOTA","Three tweaks to motion and appearance features lift MOT FCG to 63.1 HOTA","Aspect-aware GIoU and dynamic appearance lift MOT FCG to 63.1 HOTA","Improving on MOT FCG: new features yield 63.1 HOTA on MOT17","Diagonal GIoU and dynamic EMA sharpen multi-object tracking to 63.1 HOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001025,"raw_usage":{"total_tokens":4335,"prompt_tokens":974,"completion_tokens":3361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":3241}},"tokens_in":590,"tokens_out":3361,"duration_ms":22378,"temperature":1.0,"reasoning_tokens":3241,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:02:47.254928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce MOT FCG++ with the same fine-tuned YOLOX detector and SBS ReID features but swap DGIoU back to plain IoU; if HOTA on MOT17-val does not drop below the reported 76.14, the paper's attribution of the gain to the diagonal modulation fails.","supporting_citations":[{"cited_title":"Multiple Object Tracking from appearance by hierarchically clustering tracklets","cited_arxiv_id":"2210.03355","evidence_quote":"MOT FCG is the baseline hierarchical clustering tracker whose median-element appearance and IoU motion features this paper replaces."},{"cited_title":"Hybrid-sort: Weak cues matter for online multi-object tracking","cited_arxiv_id":null,"evidence_quote":"Hybrid-SORT supplies the weak-cue idea and the L2/L1 diagonal modulation term used to define DGIoU."},{"cited_title":"Observation- centric sort: Rethinking sort for robust multi-object tracking","cited_arxiv_id":null,"evidence_quote":"Deep-OC-SORT supplies the Dynamic Appearance representation with confidence-adaptive EMA weighting that the paper adapts for clustering tracklets."},{"cited_title":"Fastreid: A py- torch toolbox for general instance re-identification","cited_arxiv_id":null,"evidence_quote":"SBS (FastReID) is the network used to extract the ReID appearance embeddings."},{"cited_title":"Mot16: A bench- mark for multi-object tracking","cited_arxiv_id":null,"evidence_quote":"MOT17 is the main benchmark dataset on which the headline HOTA, MOTA, and IDF1 numbers are reported."}],"review_version":1}