{"id":"e3570e9b-b08a-4019-b6d2-1f17e1887d51","arxiv_id":"2412.06258","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A pose-based virtual marker overlay, applied to test videos, tracks 3x3 basketball players with zero ID switches and a 72.6 HOTA on a private dataset, but the markers supply identity at test time.","lead":"Sports-vmTracking overlays colored virtual markers on basketball video frames, then trains a pose model to follow each marker and converts the detected keypoints into bounding boxes. On a private 3x3 basketball dataset, the method reports a HOTA score over 10 points higher than automated trackers, but the comparison is not apples-to-apples because the test videos contain the markers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central benchmark is confounded: Sports-vmTracking receives test-time virtual markers with known player IDs while the compared trackers see raw video, so the >10-point HOTA gain and 0 ID switches likely measure the markers, not MOT quality.","rationale":"Good-faith reading: The paper proposes an annotation/training workflow, and the active-learning efficiency claim is supported by Table 1. The problem is the benchmark claims. The method's own pipeline explicitly overlays VMs on test videos, and the single-agent model is trained to use those VMs. The compared trackers do not receive this information. This is not a subtle parameter mistake; it is a mismatch in the evaluation protocol that directly explains the >10-point HOTA gap and the perfect ID score. A reviewer cannot conclude from Table 2 that the method is a better automated MOT algorithm, only that a pose model with per-frame identity hints can output consistent boxes. I agree with the reader's weakest assumption and verdict. I am not claiming fraud; the manuscript describes the protocol openly, and the open description makes the confound clear. The single concrete control experiment above would settle whether the markers matter. Until then, the central claim should not be accepted, so I recommend keeping the reader's REJECT verdict.","tokens_in":12487,"tokens_out":4691,"duration_ms":48418,"concrete_test":"Re-run the Sports-vmTracking pipeline on the same 21 test videos with the VM overlay removed or, better, with marker colors randomly permuted across players per frame so the dots no longer encode consistent identity, keeping the trained single-agent model fixed. If HOTA and ID switches remain near 72.6 and 0, the result does not depend on test-time identity cues and the concern is refuted. If HOTA falls by more than ~10 points or ID switches appear, the headline gain is an artifact of the VM input, and a fair comparison requires evaluating all methods on identical marker-free video.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 states that \"In the test video dataset with VM, we created 6 test datasets with varying sizes and quantities of VM,\" and Step 3 (Section 3.3) performs inference on these VM-labeled test videos. Section 3.2 trains the single-agent pose model on VM videos where \"ID-consistent labeling between frames [is] guided by the VMs.\" Because the overlaid VMs are colored dots placed at selected keypoints of each known player, the test input already contains per-frame identity information. Sports-vmTracking can therefore regress keypoints conditioned on marker color/location rather than solve detection or association; this is a different, easier task than the one faced by Deep-EIoU, BOT-SORT, Basketball-SORT, and maDLC (w/o VMs), which all see marker-free video. The reported 0.0 ID switches and FN=FP values are consistent with the tracker inheriting identities from the markers. The abstract's 72.3 and Table 2's 72.6 mismatch, and Section 3.3's admission that keypoint-derived boxes are \"not accurate,\" are secondary but reinforce that the headline MOT comparison is not a valid measure of automated tracking. The active-learning result (659 vs 3,817 annotated frames, 4.40 vs 4.16 px RMSE) is a reasonable separate contribution, but it does not validate the claimed MOT superiority.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Sports-vmTracking, an adaptation of the vmTracking multi-animal pose-tracking method to 3x3 basketball. The pipeline has three steps: (1) train a multi-agent DeepLabCut model to detect six keypoints per player and use these to overlay colored virtual markers (VMs) onto video frames; (2) train a single-agent pose model on the VM-augmented videos; (3) run that model on VM-augmented test videos, convert the predicted keypoints to bounding boxes, and evaluate standard MOT metrics (HOTA, DetA, AssA, FN, FP, ID switches). The authors report a HOTA of 72.6 (abstract says 72.3), about 10 points above Deep-EIoU, BOT-SORT, Basketball-SORT, and maDLC without VMs, and 0 ID switches. They also report an active-learning result in which 659 annotated frames achieve a test RMSE of 4.40 px versus 4.16 px with all 3,817 frames annotated. The central empirical claim is that the VM-based method substantially outperforms automated MOT trackers on a private 3x3 basketball dataset.","tokens_in":12731,"tokens_out":2904,"duration_ms":30023,"significance":"If the reported comparison were valid, the paper would demonstrate a practically important approach to occlusion- and appearance-challenged sports tracking, and the active-learning result would be a useful efficiency contribution for pose annotation. The authors are explicit about the pipeline and disclose that keypoint-derived bounding boxes are 'not accurate' in Section 3.3. However, the significance is undercut by a fundamental evaluation confound: Sports-vmTracking is tested on videos with identity-carrying virtual markers overlaid, while the compared automated methods are tested on raw marker-free video. The 0 ID switches and the large HOTA gap therefore reflect the input differences at least as much as any tracking capability. The active-learning comparison is a separate, plausible contribution, but it does not validate the claimed MOT superiority. The paper does not release code or data, which limits reproducibility.","major_comments":[{"comment":"The headline MOT comparison is confounded. Sports-vmTracking is evaluated on test videos onto which virtual markers with known player identities have been overlaid (Section 4.2: 'In the test video dataset with VM, we created 6 test datasets with varying sizes and quantities of VM'), and Step 3 in Section 3.3 performs inference on these VM-labeled videos. The single-agent pose model is trained on VM videos where ID-consistent labeling is 'guided by the VMs' (Section 3.2). The baselines (Deep-EIoU, BOT-SORT, Basketball-SORT, maDLC w/o VMs) are evaluated on raw marker-free video. The test inputs therefore differ in a way that directly supplies identity cues to the proposed method, so the reported 0.0 ID switches, the near-zero FN/FP, and the >10-point HOTA advantage in Table 2 do not establish that Sports-vmTracking is better at the MOT task faced by the baselines.","section":"Section 3.3, Section 4.2, Table 2"},{"comment":"The abstract reports an average HOTA of 72.3%, but Table 2 reports 72.6 ± 2.7 for Sports-vmTracking, and Table 3 reports 72.6 for the 1-pixel, 6-point configuration. The 72.3 value appears nowhere in the tables. The authors should state exactly which configuration and which number is the headline result, and reconcile the discrepancy.","section":"Abstract and Table 2"},{"comment":"The bounding-box conversion step is itself a source of major uncertainty. Section 3.3 states that the converted keypoint boxes 'are not accurate,' and Table 4 shows that when ground-truth keypoints are used, the padding method reaches HOTA 75.0, whereas the Max_Min method reaches only 43.8. This indicates that the final MOT quality is highly sensitive to the keypoint-to-box conversion and to keypoint accuracy, not merely to association. However, all reported Sports-vmTracking HOTA numbers (72.6 in Table 2) are below the ground-truth-keypoint upper bound, and the paper does not quantify how much of the 72.6 is attributable to the pose model versus the conversion method. The comparison would be much more informative if the same bounding-box conversion were applied to the keypoint outputs of the baselines, or if a marker-free variant of the proposed pipeline were reported.","section":"Section 3.3 and Table 4"},{"comment":"The active-learning result is presented as a main contribution, but it does not validate the MOT claims. Table 1 shows that 659 annotated frames yield a test RMSE of 4.40 px versus 4.16 px with 3,817 frames, which is a reasonable efficiency outcome. However, this pose model is trained and tested on VM-augmented data, and the MOT evaluation is on the same VM-augmented test set. The active-learning result would be more convincing if the pose error were reported for marker-free test frames as well, or if the downstream MOT metrics were reported as a function of annotation budget.","section":"Section 4.2, Table 1"}],"minor_comments":[{"comment":"The text uses 'Sport-vmTracking' in one paragraph ('Sport-vmTracking’s HOTA score exceeds...') while the rest of the paper uses 'Sports-vmTracking'; please unify the name.","section":"Section 4.3"},{"comment":"The table caption reads 'T able 1' with an extra space; please fix the formatting.","section":"Table 1 caption"},{"comment":"The sentence 'In vmTracking, the VM size was set to a minimum of 1 pixel' is unclear: it presumably means a single-pixel marker, but 'minimum' could be misread as a lower bound; please rephrase.","section":"Section 4.4"},{"comment":"The description of the VM creation step says 'After manually correcting some ID switches...', which is a manual intervention step, but the paper later claims an 'automated' pipeline; the role of manual correction in the test-time procedure should be clarified.","section":"Section 3.1"}],"recommendation":"reject","confidential_remarks":"The core experimental design makes the headline claim untestable as stated: the proposed method receives identity-information-rich test inputs that the baselines do not. This is not a matter of tuning or presentation; it requires either a redefinition of the task (e.g., as a marker-assisted semi-automatic tracking system) and honest comparison against analogous marker-assisted approaches, or a redesign of the evaluation to remove the test-time marker advantage. The private dataset further limits verification. The active-learning result is worth salvaging, but as a standalone contribution it is too thin for the current paper. I would not consider acceptance until the evaluation is reframed or the comparison is made fair."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: the central comparison in this paper is not a fair test of MOT. Sports-vmTracking overlays colored virtual markers with known player IDs on the test videos before running its pose model, while the automated baselines see raw video. So the 10-point HOTA gap and the 0 ID switches mostly reflect that the test input already contains the identities the tracker is supposed to infer. That is not a comparison of tracking methods.\n\nWhat is actually new: the application of vmTracking to team sports is legitimate, and the 3x3 basketball pose dataset (3,817 frames, 6 keypoints per player) could be useful to the sports-vision community. The active-learning experiment (659 vs 3,817 annotated frames, test RMSE 4.40 vs 4.16 px) is a reasonable efficiency result, and the padding-based keypoint-to-bbox conversion is a sensible engineering contribution. The VM size/quantity ablation is thorough and gives a clear takeaway: 1-pixel, 6-point VMs work best.\n\nThe soft spots are serious. The confound is load-bearing. Section 3.2 says training on VM videos uses ID-consistent labeling guided by the VMs; Section 4.2 creates test datasets with VMs; Section 3.3 runs inference on those VM-labeled videos. The baselines are evaluated without VMs. The abstract reports 72.3 HOTA while Table 2 says 72.6 ± 2.7, and Section 3.3 admits the keypoint-derived boxes are \"not accurate.\" The maDLC baseline (52.4 HOTA) is a partial nod to the VM-vs-no-VM question, but it is a different model and still leaves the main comparison asymmetric. The paper would need either a fair comparison on marker-free test video or a clear reframing as a semi-automatic annotation-assisted method, not as a replacement for automated MOT.\n\nWho this is for: readers interested in semi-automatic annotation pipelines for fixed-camera sports, or in the active-learning efficiency result, will get something. As a claim about automated MOT superiority, it does not hold.\n\nMy recommendation: I would send it to peer review rather than desk-reject. The active-learning result and dataset are worth referee time, and the confound is fixable with a reframing. But this needs major revision—either compare on marker-free video (which would likely break the method) or present it as annotation-assisted tracking and drop the over-10-points claim. A desk reject would also be defensible, but I lean toward review.","headline":"The headline HOTA gain is an artifact of test-time identity markers; the active-learning result is real but secondary.","tokens_in":13336,"tokens_out":2185,"would_cite":false,"duration_ms":22218,"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":"Overlaying pose-based virtual markers with known player identities onto 3x3 basketball video lets a single pose model track all players through heavy occlusion with zero identity switches, beating automated trackers by more than ten HOTA…","keywords":["multi-object tracking","pose estimation","virtual markers","3x3 basketball","active learning","occlusion handling","HOTA","identity switches"],"falsifier":"Run the same pose-tracking and bounding-box pipeline on the same 3x3 basketball test videos with all virtual-marker overlays removed; if HOTA falls to the level of the automated baselines, the reported gain and zero ID switches come from the injected identity cues rather than from the tracking architecture.","tokens_in":12219,"feed_emoji":"🏀","tokens_out":11444,"duration_ms":93636,"temperature":0.7,"pith_summary":"This paper tries to establish that in team-sports tracking, the hard problem of keeping players' identities straight under occlusion and near-identical uniforms can be bypassed by painting virtual markers—colored keypoints with known IDs—onto the video, and then tracking those markers. On the authors' 3x3 basketball dataset, the method reports a HOTA score (a tracking metric balancing detection and association accuracy) of 72.3%, more than ten points above recent fully automated trackers, with zero identity switches; the results table lists 72.6 ± 2.7. The paper also shows that active learning on only 659 of 3,817 frames yields pose accuracy nearly matching full annotation (4.40 vs 4.16 pixels test error), cutting labeling cost. The central claim is that pose-based virtual markers, not learned appearance association, are the effective mechanism for robust multi-object tracking in crowded, similar-appearance scenes.","feed_headline":"Virtual markers deliver zero ID switches, 10+ HOTA in 3x3 basketball","feed_subtitle":"Overlaid keypoints with known identities beat appearance-based trackers and cut annotation cost via active learning.","key_machinery":"The central object is the virtual marker (VM): a colored dot or set of dots overlaid on video frames at known player keypoints, carrying identity through color. The pipeline has three steps: a multi-agent pose estimator is trained with active learning to generate VMs; those VMs are overlaid on training videos to force ID-consistent annotation and on test videos to give the tracker identity cues; then a single-agent pose estimator outputs six keypoints per player, which are converted to bounding boxes with a padding offset. The VMs do the association work that appearance-based re-identification cannot: they make 'who is who' visible in the image, leaving the tracker with the simpler task of precise keypoint localization rather than identity reasoning.","core_discovery":"The central claim is that a pose-based virtual-marker pipeline can solve multi-object tracking in team sports under heavy occlusion and appearance similarity. The method trains a multi-agent pose estimator to generate colored keypoint markers with known player identities, overlays those markers on both training and test videos, and then tracks each player's six keypoints with a single-agent pose estimator; the keypoints are converted into bounding boxes for comparison with automated MOT methods. On the authors' 3x3 basketball dataset, this pipeline achieves a HOTA of 72.3% (72.6 ± 2.7 in Table 2), more than ten points above marker-free baselines, with zero identity switches. The authors attribute the gain to the markers providing a direct visual identity cue, which keeps association robust when players occlude each other or wear identical uniforms, while active learning reduces the annotation burden.","pith_inferences":["If overlaying identity cues at test time is accepted, the method reframes sports MOT as a marker-tracking problem: the hard part shifts from data association to precise keypoint localization, so improvements in pose stability translate directly into tracking scores.","The paper leaves open whether the same advantage would appear on public benchmarks, which do not provide pre-assigned player identities to overlay; a test on a public dataset with markers generated from its own pose annotations would indicate generalizability.","The zero-ID-switch result is partly a consequence of the markers being visible in the input: if the pose model can read the marker color, identity is supplied by the image rather than inferred, so the metric may be measuring marker-visibility fidelity as much as tracking ability.","Because active learning selected occlusion-heavy frames for annotation, the approach could adapt to other fixed-camera team sports such as futsal or handball, where players are few, uniforms are similar, and occlusions cluster around the ball."],"forward_implications":["If the reported results hold, adding pose-based virtual markers to team-sports video raises HOTA by more than ten points over fully automated trackers and eliminates identity switches in 3x3 basketball.","Active learning on occlusion-heavy frames cuts the annotated training set by roughly a factor of six (659 vs 3,817 frames) while keeping test error within 0.24 pixels of full annotation.","Smaller markers work better: 1-pixel VMs outperform 3-pixel VMs at every marker count, and six VMs per player outperform one or three, so dense, unobtrusive identity cues are the recommended configuration.","The way keypoints are converted to bounding boxes matters as much as the tracking itself: the padding method scores HOTA 75.0 versus 43.8 for the max-min method, so the conversion step is a major component of the reported gains."],"supporting_citations":[{"why":"Introduces virtual markers for multi-animal pose tracking, the base method this paper extends to team sports.","marker":"[18]"},{"why":"Provides the open-source pose-estimation toolbox and active-learning workflow used in both model stages.","marker":"[39]"},{"why":"Supplies the multi-agent pose model used to generate virtual markers from annotated frames.","marker":"[32]"},{"why":"Supplies the single-agent backbone that tracks virtual markers once they are overlaid on video.","marker":"[41]"},{"why":"Defines the HOTA metric and its DetA/AssA/LocA components used for all reported comparisons.","marker":"[21]"},{"why":"Provides the raw 3x3 basketball video data and participant consent underlying the new dataset.","marker":"[19]"},{"why":"One of the fully automated MOT baselines compared on the same test videos.","marker":"[4]"},{"why":"A basketball-specific automated MOT baseline used to benchmark the VM method.","marker":"[8]"},{"why":"A sports MOT baseline using expanded IoU association, compared without virtual markers.","marker":"[9]"},{"why":"The public basketball dataset whose annotation scope and volume the paper compares against its own dataset.","marker":"[20]"}],"fun_headline_variants":["Pose markers end ID switches in 3x3 basketball","Virtual keypoints outrun appearance trackers by 10 HOTA","Zero ID switches with pose-based virtual markers in hoops","Active learning slashes annotation, lifts HOTA to 72.3","Occlusion-proof: virtual markers track basketball players"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that overlaying virtual markers with known player IDs onto the test videos is a legitimate part of the tracking pipeline, so the comparison with fully automated trackers that do not receive these identity cues measures the same task.","fun_headline_variants_meta":{"raw":{"variants":["Pose markers end ID switches in 3x3 basketball","Virtual keypoints outrun appearance trackers by 10 HOTA","Zero ID switches with pose-based virtual markers in hoops","Active learning slashes annotation, lifts HOTA to 72.3","Occlusion-proof: virtual markers track basketball players"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":3023,"prompt_tokens":991,"completion_tokens":2032,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1946}},"tokens_in":607,"tokens_out":2032,"duration_ms":17132,"temperature":1.0,"reasoning_tokens":1946,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:51:06.455501+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pose-tracking and bounding-box pipeline on the same 3x3 basketball test videos with all virtual-marker overlays removed; if HOTA falls to the level of the automated baselines, the reported gain and zero ID switches come from the injected identity cues rather than from the tracking architecture.","supporting_citations":[{"cited_title":"bioRxiv (2024) https://doi.org/10.1101/2024.02.07.579241 https://www.biorxiv.org/content/early/2024/11/30/2024.02.07.579241.full.pdf","cited_arxiv_id":null,"evidence_quote":"Introduces virtual markers for multi-animal pose tracking, the base method this paper extends to team sports."},{"cited_title":"Nature Methods19(4), 496–504 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-agent pose model used to generate virtual markers from annotated frames."},{"cited_title":"Multimedia Tools and Applications, 1–17 (2024)","cited_arxiv_id":null,"evidence_quote":"A basketball-specific automated MOT baseline used to benchmark the VM method."},{"cited_title":"In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"A sports MOT baseline using expanded IoU association, compared without virtual markers."},{"cited_title":"In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports, pp","cited_arxiv_id":null,"evidence_quote":"The public basketball dataset whose annotation scope and volume the paper compares against its own dataset."}],"review_version":1}