{"id":"87782c93-5f3b-4522-98a9-78127041274c","arxiv_id":"2506.00774","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free MOT framework that adds zero-shot depth histograms and a hierarchical box/mask alignment score to association, with mixed state-of-the-art results.","lead":"This paper presents a video tracker that measures each object's distance with a pretrained depth model and combines that with a shape-matching score to follow players across occlusions and crossings. The method reports top marks on a sports-tracking benchmark for some metrics, but on another benchmark it only ties a trained rival and scores lower on detection-related metrics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7) sums four score types with heterogeneous ranges and no normalization, so the reported gains may depend on an unstated implicit weighting that the 'no fine-tuning' claim does not account for.","rationale":"The reader's weakest assumption concerns depth-histogram stability; I agree that is fragile, but the more load-bearing gap is upstream: even a perfect depth histogram is inserted into an unnormalized sum. Because Eq. (7) adds SHAS in [0, e] to cosine terms in [-1, 1], the effective weight of each cue is determined by the arbitrary scales of the raw scores. The paper does not specify normalization or weights, so the reported rankings are not reproducible, and the 'no training nor fine-tuning' claim hides implicit calibration. A concrete scaling test would settle whether the concern lands. If the test shows sensitivity, the paper must disclose or fix score normalization; if it shows invariance, the concern is resolved. This does not alter the reader's CONDITIONAL verdict; it reinforces it. I deliberately did not base the attack on the 'state-of-the-art' overclaim (Table 1 shows MOTRv2 HOTA 69.9 > DepthMOT 64.27), nor on the novelty concern with [16], because the score-fusion issue is more fundamental and more directly testable.","tokens_in":828,"tokens_out":1067,"duration_ms":81712,"concrete_test":"Re-run the DanceTrack validation protocol of Table 3 with Sdepth scaled by factors {0.1, 0.5, 2.0, 5.0, 10.0}, keeping all other terms fixed; also re-run with each of the four score matrices z-scored across the current cost matrix before summation. If HOTA changes by more than about 1 point, or if the marginal benefit of depth changes sign, then the claimed improvement is an artifact of the unstated relative scale and the method is under-specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central association step (Eq. 7) is Match_t = SHAS + Sang + Sdepth + Semb, with no weights or normalization. SHAS, from Eq. 6, equals SIoUbbox * exp(SIoUSeg) and lies in [0, e] (~2.72), while the other terms are cosine similarities typically in [-1, 1] or [0, 1]. Thus the HAS term can dominate the sum by a factor of about 2-3, and the relative influence of depth and appearance is set by the arbitrary ranges of raw scores. The paper never reports score normalization, clipping, or weights, nor any sensitivity analysis. Because the central claim is that DepthMOT improves association 'without any training nor fine-tuning', this unstated calibration is a hidden hyperparameter: different histogram bin counts, feature normalizations, or a simple z-score would change the implicit weighting and potentially the reported rankings (e.g., DanceTrack HOTA 64.27 vs DiffMOT 62.3). Table 3 is too under-specified to reveal the magnitude or direction of this scale dependence, and it also lacks a clean baseline without both HAS and depth. The reported numbers are therefore not reproducible from the text alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"DepthMOT is a tracking-by-detection MOT framework that augments association with three cues on top of a non-linear Kalman filter with OC-SORT-style ORU: zero-shot monocular depth (Depth Pro), a Hierarchical Alignment Score (HAS) combining bounding-box IoU with SAM2 mask IoU via SHAS = SIoU_bbox * exp(SIoU_seg), and FastReID appearance embeddings with EMA updating. The total matching score (Eq. 7) is an unweighted sum of the HAS, angle, depth, and appearance components, solved with a linear assignment. Depth and appearance are applied only to high-confidence detections, while low-confidence detections are associated by IoU alone in a second stage. The paper reports gains over DiffMOT on DanceTrack (HOTA 64.27 vs 62.3, AssA 51.04 vs 47.2, IDF1 66.54 vs 63.0) and comparable results on SportsMOT, with MOT17/MOT20 results deferred to an external supplementary. An ablation on the DanceTrack validation set (Table 3) attributes the gains to HAS and depth.","tokens_in":10389,"tokens_out":12500,"duration_ms":112978,"significance":"Conditional significance is real but moderate. If the associations hold, the paper provides a clean, training-free depth augmentation that improves association metrics (HOTA/IDF1/AssA) on the non-linear-motion DanceTrack benchmark, a setting where IoU-only association is known to fail; the release of code and the use of the same YOLOX detector as DiffMOT make the comparison concrete and falsifiable. The paper also makes an explicit, checkable novelty claim (first use of monocular depth as an independent decision matrix) and is honest in the text that no single tracker wins on MOT17/MOT20. However, the gains are selective: MOTA and DetA drop on both benchmarks, the MOT17/MOT20 evidence is outside the manuscript, and the reproducibility gaps in the depth-histogram specification and in the calibration of Eq. (7) currently prevent verification of the central attribution of the gains to depth.","major_comments":[{"comment":"The total matching score sums four components with heterogeneous ranges: SHAS lies in [0, e] (since SIoU_bbox <= 1 and exp(SIoU_seg) <= e), while Sang, Sdepth, and Semb are cosine similarities in [-1, 1] or [0, 1]. The equal-weight sum therefore contains an implicit calibration: at high bounding-box overlap the HAS term can outweigh depth and appearance by a factor of two or more, and any change in histogram binning, embedding normalization, or a simple z-score would alter the relative contribution of the depth channel. Since the paper's headline claim is 'without any training nor fine-tuning', the authors must either report normalized components and their weights, or provide a sensitivity analysis showing that the Table 1 and Table 3 rankings are stable under renormalization of the four score types.","section":"Sec. 3.5, Eq. (7)"},{"comment":"The depth score is never specified operationally: the paper does not state the number of histogram bins, the binning strategy (fixed range vs per-frame percentile), the normalization (count, density, or L2), how background pixels inside the bounding box are masked out, or how Depth Pro's per-scene scale ambiguity is handled across frames. Because the depth histogram is the single new cue claimed to drive the association gains, this omission makes the central experiments unreproducible from the text; the statement that histograms 'yielded the most effective results' also implies a selection process that should be reported as part of the 'no fine-tuning' claim.","section":"Sec. 3.3"},{"comment":"The ablation table is unreadable as typeset: the checkmarks do not align with the column headers, the header row is merged ('DepthHOTA'), and the rows are concatenated, so a reader cannot determine which configuration produced 61.81 HOTA. More substantively, the table has no row that excludes both HAS and depth, and the apparent sequences of checkmarks suggest configurations whose comparison (e.g., 60.62 vs 60.45 vs 61.81) would need pairwise toggles of depth and HAS to attribute the gains; as printed, the core claim that 'the combination of both HAS and depth yields the highest accuracy' cannot be audited.","section":"Table 3"},{"comment":"The text states that DepthMOT 'outperforms' DiffMOT on DanceTrack, but DepthMOT is lower on MOTA (90.08 vs 92.8) and DetA (81.07 vs 82.5), and on SportsMOT it ties DiffMOT on HOTA (76.2) while losing on MOTA (95.9 vs 97.1) and DetA (88.5 vs 89.3). Since the abstract claims 'state-of-the-art results on challenging benchmarks', the systematic detection-side regression (possibly from the low-confidence second-stage association, which is pure IoU, or from depth failing on small or heavily occluded boxes) must be analyzed, and the claim should be qualified to association metrics rather than left implicit.","section":"Tables 1 and 2; Sec. 4.1"},{"comment":"The Fig. 2 caption states that the exponential term amplifies fine-grained segmentation 'particularly when bounding box overlap is low', but SHAS = SIoU_bbox * exp(SIoU_seg) vanishes as SIoU_bbox approaches 0, so pixel-level shape similarity cannot rescue a match when the bounding boxes do not overlap. The hierarchical narrative (shape influence grows as coarse alignment improves) is consistent with the product form only in the large-overlap regime; the caption and the text must be reconciled with the formula, or the formula changed, since this score is a claimed contribution.","section":"Sec. 3.4, Eq. (6) and Fig. 2 caption"},{"comment":"The breadth of the benchmark claim rests on a non-archival external link (sigport.org), and the sentence 'Ours achieves the lowest false positive rate' has no locale in the main text: it is unclear whether it refers to MOT17, MOT20, or both, and the supporting numbers are absent. The manuscript should include the MOT17/MOT20 tables (or cite a permanently archived supplementary) and state the scope of the claim; note that the paper itself concedes 'under MOT17 and MOT20 datasets (linear motion), no tracker performs best overall', which is a limitation that the abstract's unconditional phrasing does not reflect.","section":"Sec. 4.1, MOT17/MOT20"}],"minor_comments":[{"comment":"State the range of the dynamic coefficient C explicitly: with T=0.95 and a confidence threshold of 0.6, C lies in [0.95, 1], so a new embedding receives weight at most 0.05; this near-frozen update deserves a sentence of justification and a check against the Deep OC-SORT reference.","section":"Sec. 3.1, Eq. (5)"},{"comment":"The angle score Sang is referenced in Eq. (7) but never defined in the paper; give its formula (e.g., the direction cosine between tracklet velocity and observation displacement used in Deep OC-SORT).","section":"Sec. 3.5 and Eq. (7)"},{"comment":"The novelty claim ('first MOT framework to incorporate 3D features (monocular depth) as an independent decision matrix') should be sharpened against Quach et al. [16] and Wang et al. [17], who already inject depth into the association step via a Kalman-filter state and a stereo-aware OCSORT variant; the distinction needs to be argued in the main text, not only asserted through the phrase 'independent decision matrix'.","section":"Sec. 2 and Sec. 3.3"},{"comment":"No runtime or hardware details are reported; since the paper argues an efficiency advantage over MOTRv2, report frames-per-second and the per-frame cost of Depth Pro and SAM2.","section":"Sec. 4"},{"comment":"The ablation is performed on the validation set and the final configuration is selected there; disclose this selection explicitly so the 'no fine-tuning' claim is unambiguous about validation-based design choices.","section":"Sec. 4.2"},{"comment":"Minor presentation items: the blue/bold highlighting convention described in Sec. 4 is not visible in the typeset tables, and the phrase 'without any training nor fine-tuning' should be 'without any training or fine-tuning'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main editorial risk is novelty disclosure: the phrase 'first MOT framework to incorporate 3D features (monocular depth) as an independent decision matrix in the association step' is narrow enough to be defensible, but Quach et al. [16] already use relative depth ordering inside the Kalman update for association, so the authors should be pushed to articulate what 'independent decision matrix' adds beyond [16] before acceptance. The supplementary is hosted on sigport.org, which is not a permanent archive; the editor may want to require the MOT17/MOT20 numbers to be included in the paper. The metric-selective reporting (leading on HOTA/IDF1/AssA, trailing on MOTA/DetA on both benchmarks) should be copy-edited so the abstract and Sec. 4.1 match the tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a training-free, plug-in association score for MOT: zero-shot depth histograms plus a product-of-IoUs term (box IoU × exp(mask IoU)). The combination is genuinely new as far as I can tell, and the authors ship code and test on two occlusion-heavy benchmarks. On DanceTrack they beat DiffMOT on HOTA/IDF1/AssA; on SportsMOT they tie on HOTA with better IDF1/AssA. That is credible evidence for a small contribution.\n\nThe soft spots are real. Eq. (7) adds four scores with different ranges. The HAS term lives roughly in [0, e] while the other three are cosine similarities in [-1,1] or [0,1], so without any normalization HAS implicitly dominates by a factor of two or three. The paper gives no weights, no clipping, no z-scoring, no sensitivity analysis. That is a hidden calibration hyperparameter behind a 'no training/fine-tuning' claim. The authors' own tables hint at this: they win on association metrics but lose on MOTA/DetA on both benchmarks, which is what you'd expect if shape/depth cues crowd out box overlap.\n\nAlso missing: histogram bin count and normalization in the depth descriptor, and any description of how they handle depth map noise or partial occlusion. The ablation table is too under-specified to parse cleanly. MOT17/MOT20 results are deferred to an external supplementary link, which is a big hole given the abstract's broad claim. And the 'first MOT framework to incorporate depth as an independent decision matrix' phrasing does not survive contact with their own reference [16], which updates Kalman states with depth ordering. The histogram formulation is new; the priority claim is too strong.\n\nNone of this is fatal. The core idea is simple, sensible, and mostly new; the evidence is partial but real. The fixes are easy: report score normalization and bin counts, add a clean ablation with an unmodified baseline, and pull MOT17/MOT20 into the paper. I'd send it to review and ask for those specifics before acceptance. A MOT researcher gets real value from the depth-histogram cue and the HAS formulation. I wouldn't cite it yet until the calibration question is answered.","headline":"Genuinely new training-free depth+MOT scoring with credible partial results, undercut by missing normalization details and deferred benchmark numbers.","tokens_in":10996,"tokens_out":3804,"would_cite":false,"duration_ms":32489,"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":"Depth-aware, training-free tracker beats trained rivals on DanceTrack","keywords":["multiple object tracking","monocular depth estimation","zero-shot depth","Hierarchical Alignment Score","data association","tracking-by-detection","pixel-level alignment","DepthMOT"],"falsifier":"Construct a sequence of two visually identical objects that cross paths while remaining at nearly the same distance from the camera; if the depth-histogram cosine similarity between the two different objects is comparable to the same-object similarity and the mask IoU also matches both boxes, the tracker should produce an identity switch at the crossing. Alternatively, recompute the reported DanceTrack results while varying the depth-histogram bin count; if HOTA changes by more than a few points, the claimed gains depend on an unspecified implementation detail.","tokens_in":9926,"feed_emoji":"🎯","tokens_out":4465,"duration_ms":37237,"temperature":0.7,"pith_summary":"DepthMOT aims to show that a multiple-object tracker can improve association accuracy without any training by adding two cheap cues to the standard bounding-box overlap score: monocular depth and pixel-level shape alignment. The paper argues that motion- and appearance-based scores alone confuse objects that look alike or overlap, whereas a depth histogram of each detection and a mask-level IoU between tracklet and detection disambiguate them. It reports state-of-the-art results on DanceTrack, improving HOTA by 1.97 points, AssA by 3.84 points, and IDF1 by 3.54 points over the trained DiffMOT baseline, and matching DiffMOT on SportsMOT. If correct, the framework would give a plug-in, training-free route to better tracking in crowded and occluded scenes.","feed_headline":"Depth-aware, training-free tracker beats trained rivals on DanceTrack","feed_subtitle":"Monocular depth histograms and pixel-level mask overlap lift association accuracy in crowded scenes.","key_machinery":"The load-bearing objects are the per-detection depth histogram and the Hierarchical Alignment Score (HAS). Zero-shot depth estimation (Depth Pro) produces a dense relative depth map per frame; for each bounding box the depth values are collected into a histogram, and cosine similarity between histograms of a tracklet and a new detection becomes the depth score $S_{\\mathrm{depth}}$. HAS combines bounding-box IoU with segmentation-mask IoU from a promptable video segmentation model (SAM2), using the multiplicative-exponential form $S_{\\mathrm{HAS}} = S_{\\mathrm{IoU}}^{\\mathrm{bbox}} \\exp(S_{\\mathrm{IoU}}^{\\mathrm{seg}})$ so that fine-grained shape alignment matters most once boxes already overlap. The final matching cost sums HAS, directional motion similarity, depth similarity, and appearance similarity from a pretrained re-identification model, and is passed to a linear assignment solver.","core_discovery":"The paper's central claim is that a training-free association process can match or beat trained trackers by treating monocular depth as an independent decision matrix and by refining IoU with a Hierarchical Alignment Score. The depth score is obtained by running a zero-shot depth estimator on each frame, building a histogram of depth values inside every detection box, and computing cosine similarity between histograms across frames. The Hierarchical Alignment Score is defined as $S_{\\mathrm{HAS}}(\\hat{X},D) = S_{\\mathrm{IoU}}^{\\mathrm{bbox}}(\\hat{X},D) \\times \\exp\\bigl(S_{\\mathrm{IoU}}^{\\mathrm{seg}}(\\hat{X},D)\\bigr)$, so coarse bounding-box overlap dominates at first and pixel-level mask overlap, produced by a promptable video segmentation model, sharpens the score as alignment improves. These are added to directional motion and appearance scores, and the total matrix is solved as a linear assignment. On DanceTrack the paper reports HOTA 64.27, IDF1 66.54, and AssA 51.04, surpassing the trained DiffMOT tracker; on SportsMOT it reports HOTA 76.2, comparable to DiffMOT. The authors also claim this is the first MOT framework to use monocular depth as an independent association feature.","pith_inferences":["A testable extension is to swap Depth Pro for any metric-depth estimator and check whether the histogram descriptor remains discriminative; the paper's argument implies the cue, not the specific model, is what matters.","One implicit consequence is that the depth histogram could serve as a weak ReID signature in appearance-free regimes, e.g., low-resolution or grayscale footage where appearance models degrade.","The method may inherit the limitations of the segmentation model: if the promptable segmentation fails on heavily truncated objects, the exponential mask term could mislead rather than disambiguate.","It would be worth measuring performance with the depth histogram's bin count and normalization varied, since the preprint does not fix or analyze those choices."],"forward_implications":["Depth as an independent cue should reduce identity switches when two objects have similar appearance but stand at different distances from the camera.","The HAS score should make association more robust to partial occlusion, because mask IoU can still match when bounding boxes overlap heavily.","Because no component is trained or fine-tuned, the framework can be dropped into existing tracking-by-detection pipelines without per-dataset training.","On the reported benchmarks, the gap over DiffMOT on DanceTrack suggests the largest gains appear under non-linear motion and frequent occlusion rather than linear pedestrian scenes.","The ablation indicates that depth and HAS contribute complementary gains: HAS alone raises HOTA from 54.78 to 60.62, and adding depth raises it further to 61.81 on the DanceTrack validation set."],"supporting_citations":[{"why":"Depth Pro supplies the zero-shot metric depth maps from which per-box depth histograms are built.","marker":"[27]"},{"why":"SAM2 provides the promptable video segmentation masks used to compute the pixel-level IoU in HAS.","marker":"[24]"},{"why":"OC-SORT contributes the observation-centric Kalman filter and re-update used for motion prediction and backtracking.","marker":"[6]"},{"why":"Deep OC-SORT supplies the appearance embedding update rule and directional similarity formulation that DepthMOT reuses.","marker":"[7]"},{"why":"ByteTrack provides the confidence-based division of detections into high- and low-confidence groups for the two-step association.","marker":"[19]"},{"why":"DiffMOT is the trained state-of-the-art baseline that DepthMOT compares against and claims to surpass on DanceTrack.","marker":"[2]"},{"why":"HOTA defines the primary evaluation metrics (HOTA, AssA, IDF1) used to measure association accuracy.","marker":"[29]"},{"why":"YOLOX is the object detector that produces the detections the tracker associates.","marker":"[13]"}],"fun_headline_variants":["Depth-aware MOT beats trained rivals without any training","Zero-shot depth and pixel alignment lift MOT tracking accuracy","First MOT to use depth as independent feature, no training needed","Depth histograms and mask overlap refine object association in MOT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the zero-shot depth histogram inside a detection box is a stable and distinctive signature of that object across frames, despite changes in scale, partial occlusion, and depth-map noise.","fun_headline_variants_meta":{"raw":{"variants":["Depth-aware MOT beats trained rivals without any training","Zero-shot depth and pixel alignment lift MOT tracking accuracy","First MOT to use depth as independent feature, no training needed","Depth histograms and mask overlap refine object association in MOT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3050,"prompt_tokens":965,"completion_tokens":2085,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":2020}},"tokens_in":581,"tokens_out":2085,"duration_ms":13014,"temperature":1.0,"reasoning_tokens":2020,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:58:00.344687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a sequence of two visually identical objects that cross paths while remaining at nearly the same distance from the camera; if the depth-histogram cosine similarity between the two different objects is comparable to the same-object similarity and the mask IoU also matches both boxes, the tracker should produce an identity switch at the crossing. Alternatively, recompute the reported DanceTrack results while varying the depth-histogram bin count; if HOTA changes by more than a few points, the claimed gains depend on an unspecified implementation detail.","supporting_citations":[{"cited_title":"DET: Depth-Enhanced Tracker to Mitigate Severe Occlusion and Homogeneous Appearance Problems for In- door Multiple-Object Tracking,","cited_arxiv_id":null,"evidence_quote":"SAM2 provides the promptable video segmentation masks used to compute the pixel-level IoU in HAS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"OC-SORT contributes the observation-centric Kalman filter and re-update used for motion prediction and backtracking."},{"cited_title":"Lost and Found: Overcoming Detector Failures in Online Multi-object Tracking,","cited_arxiv_id":null,"evidence_quote":"Deep OC-SORT supplies the appearance embedding update rule and directional similarity formulation that DepthMOT reuses."},{"cited_title":"Depth-Aware Scoring and Hierarchical Alignment for Multiple Object Tracking","cited_arxiv_id":"2506.00774","evidence_quote":"DiffMOT is the trained state-of-the-art baseline that DepthMOT compares against and claims to surpass on DanceTrack."},{"cited_title":"Learning the What and How of Annotation in Video Object Segmentation,","cited_arxiv_id":null,"evidence_quote":"HOTA defines the primary evaluation metrics (HOTA, AssA, IDF1) used to measure association accuracy."},{"cited_title":"Deep OC-SORT: Multi-Pedestrian Tracking by Adaptive Re-Identification,","cited_arxiv_id":null,"evidence_quote":"YOLOX is the object detector that produces the detections the tracker associates."}],"review_version":1}