{"id":"b8978481-2e06-46c2-9565-a9ca676a7963","arxiv_id":"2505.17201","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A YOLOv8-ByteTrack pipeline plus stereo triangulation can produce 3D fish tracks for some underwater video pairs, but the claimed multi-view accuracy improvement is not demonstrated.","lead":"The paper adapts YOLOv8 and ByteTrack to detect and track fish in paired stereo underwater videos, then uses epipolar geometry to triangulate 3D positions. It reports that the single-view tracker only detects about 47 percent of fish within a margin and that stereo matching works well only on long videos.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed tracking-accuracy enhancement is never measured: the paper reports no multi-view vs single-view MOT comparison, and the pipeline only adds triangulated 3D coordinates to unchanged single-view tracks.","rationale":"The paper honestly reports weak single-view tracking and stereo matching failures, which is to its credit. But the central claim in the abstract and Section 6 is that the multi-view framework enhances tracking accuracy; no experiment compares multi-view tracking accuracy against the single-view baseline. The framework's architecture cannot improve single-view tracking because it never uses cross-view information to correct detections or associations; it only triangulates existing IDs. The stereo-matching fragility (22-36% on short videos, manual re-matching) further limits any practical enhancement. Therefore, the claim is unsupported. This mirrors the reader's overall REJECT verdict, though the identifying weak link differs: the reader emphasized stereo correspondence, whereas the more fundamental issue is the absence of any accuracy comparison and the architectural impossibility of the claimed improvement. A reframed paper that presents the 3D output purely as a visualization tool could be acceptable, but as written the central claim fails.","tokens_in":16082,"tokens_out":4340,"duration_ms":43285,"concrete_test":"Re-run the evaluation on video pairs 129 and 406 where stereo matching was >95%: compute HOTA, MOTA, and IDF1 on the final multi-view output and compare with the single-view output on the same videos. Also diff the track IDs and bounding boxes between the multi-view and single-view outputs to confirm whether the pipeline changes any 2D tracking decision; if the multi-view output is identical to single-view output except for an added z coordinate, the claimed 'enhanced tracking accuracy' is not achievable by the framework's design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6 concludes that the proposed multi-view MOT framework 'enhanc[es] fish tracking accuracy.' No evaluated metric supports this. Table 1 and Appendix F report single-view MOT metrics (HOTA 0.100, MOTA 0.506, etc.); Table 6 reports stereo-matching correctness; Table 5 and Figure 6 show 3D coordinates but contain no accuracy assessment. The pipeline in Section 3.1.4 runs YOLOv8+ByteTrack independently per view, then performs stereo ID matching (step 4) and triangulation (step 6); nothing feeds cross-view information back to correct detections, identity switches, or missed tracks. Therefore the multi-view output is the same single-view 2D tracks with a z-coordinate appended. Even with perfect stereo matching, tracking accuracy as measured by HOTA/MOTA/IDF1 would be unchanged. Section 4.5 further reports that short-video matches were only 22-36% correct and had to be re-matched manually, so the claimed enhancement fails on those videos. The conclusion asserts a benefit that the experimental design cannot detect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-view multiple-object tracking framework for underwater fish videos. The pipeline runs YOLOv8 with ByteTrack independently on each stereo view, applies a post-hoc ID reconciliation heuristic, matches fish identities across views via epipolar geometry, triangulates matched bounding-box centers to 3D coordinates, and produces 3D trajectory visualizations. The authors report single-view MOT metrics, detection-margin statistics, ID-count changes after re-identification, stereo-matching accuracy, and example 3D coordinates. The central claim is that this framework enhances fish tracking accuracy for ecological studies by leveraging stereo inputs and generating a new 3D output.","tokens_in":16339,"tokens_out":4907,"duration_ms":54313,"significance":"If the enhancement claim were supported, the paper would offer ecologists a practical recipe for obtaining 3D fish trajectories from calibrated stereo cameras using off-the-shelf single-view trackers. The authors document a complete pipeline, make code available, and test on a real underwater fish dataset, which are useful engineering contributions. However, the load-bearing assertion that multi-view processing improves tracking accuracy is never tested: the pipeline adds no cross-view feedback to the single-view tracks, and no multi-view versus single-view comparison is reported. The reported metrics are also weak (recall 0.48, HOTA 0.100, DetA 0.023), and the stereo-correspondence step succeeds on only 22-36% of frames for short videos, requiring manual re-matching. As it stands, the contribution reduces to appending triangulated 3D coordinates to independently computed 2D tracks, so the significance of the claimed enhancement is not established.","major_comments":[{"comment":"The central claim that the multi-view framework enhances tracking accuracy is never tested. In the pipeline of Section 3.1.4, YOLOv8+ByteTrack is run independently on each view (step 1), and the stereo matching (step 4) and triangulation (step 6) operate only on the resulting track files; no cross-view information is fed back to correct detections, associations, or missed tracks. Consequently, the 2D tracking metrics (HOTA, MOTA, IDF1) are identical with and without the multi-view stages. The paper reports no multi-view versus single-view comparison and no metric that measures the claimed enhancement. The statement in Section 6 that the framework 'thereby enhancing fish tracking accuracy' is therefore unsupported by the experimental design.","section":"Section 6, Section 3.1.4"},{"comment":"The stereo-correspondence step, which is essential for the 3D output, is correct in only 22% and 36% of frames for the short videos, and the text states that those matches 'had to be re-matched manually' to be usable. Thus, for the short-video subset the 3D coordinates are not produced by the proposed framework but by manual annotation. The paper does not report how many of the videos and frames used in the 3D evaluation were manually re-matched, so the claimed automatic generation of 3D trajectories is established only for the two long videos.","section":"Section 4.5, Table 6"},{"comment":"The evaluation relies on an ID mapping function asserted to have '100% accuracy' in mapping tracking IDs to ground-truth IDs, but the paper gives no description of how this mapping is obtained or validated. If this mapping is imperfect, every association-sensitive metric (HOTA, AssA, IDF1, MOTA) and the re-identification evaluation in Section 4.3 are affected. This assertion is load-bearing and must be substantiated.","section":"Section 3.2"},{"comment":"The post-track re-identification heuristic (Section 3.1.4 step 2) is not validated. Table 4 shows that after re-identification the number of IDs drops below the ground-truth number for several videos (e.g., video 13_1: GT 13, before re-ID 17, after re-ID 6; video 8_1: GT 9, after re-ID 6), indicating that distinct fish are being merged. The thresholds (50 pixels, 100 frames, 10 frames, 30 frames) are introduced without ablation or sensitivity analysis, so the claimed improvement from re-identification is not established.","section":"Section 4.3, Table 4"}],"minor_comments":[{"comment":"The phrase 'novel 3D output' overstates standard triangulation from calibrated stereo cameras; suggest rephrasing to '3D output' or similar.","section":"Abstract"},{"comment":"The text alternates between 'keypoints' (e.g., corners) and fish bounding-box centers; clarify what features are actually matched and whether fish centers are used or arbitrary corner features.","section":"Section 3.1.4, step 4"},{"comment":"The 'Average*' row is not defined (which videos are included, how many), and no standard deviations are reported; add the full protocol and variance.","section":"Section 4.1, Table 1"},{"comment":"The margin criterion (half the average bounding box size and twice the average size) needs a precise definition of which average is used and how the two tiers are combined; currently the reported percentages are difficult to interpret.","section":"Section 4.2, Table 3"},{"comment":"The post-hoc heuristic is called 're-identification,' which may be confused with the learned re-ID branch of MOT models; rename it (e.g., 'ID reconciliation') to avoid ambiguity.","section":"Section 4.3 and Section 3.1.4 step 2"},{"comment":"The model trained by Dr. Nejadasl is described only as having better performance; provide training data, hyperparameters, and validation metrics so the experiments are reproducible.","section":"Section 3.1.3"},{"comment":"Several figure captions in Appendix C appear mis-numbered (multiple figures labeled 'Figure 12'), which should be corrected.","section":"Appendix C"}],"recommendation":"reject","confidential_remarks":"The paper is more a technical report or thesis chapter than a finished research article. The main claim is untested by the experimental design, and the pipeline structure makes it impossible for the multi-view stages to improve the evaluated 2D tracking metrics. If the authors resubmit a version that (a) drops or substantially redefines the enhancement claim, (b) provides a genuine multi-view versus single-view comparison, and (c) validates the stereo-correspondence and ID-mapping steps, the work could be reconsidered in a venue appropriate for applied systems papers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid engineering write-up of a YOLOv8+ByteTrack stereo pipeline for fish, but the load-bearing claim—that the multi-view framework improves tracking accuracy—is not measured. The multi-view part runs the single-view trackers independently, then matches IDs and triangulates; nothing is fed back to correct detections or associations. So HOTA/MOTA/IDF1 would be unchanged by construction. The paper's own results support this: Table 1 shows single-view metrics, Table 6 shows stereo-matching accuracy, and nowhere is there a multi-view versus single-view comparison on the same videos. The conclusion in Section 6 that the framework 'enhanc[es] fish tracking accuracy' is asserted, not demonstrated. The stress-test note lands.\n\nWhat is actually new: the underwater stereo fish dataset (182 videos, 84,464 frames, 3,796 tracks) and the detailed pipeline description are useful. The paper reports MOT metrics honestly—HOTA 0.100, DetA 0.023, recall 0.48—and compares against MOT17, which is good practice. The post-track re-ID heuristic cuts ID count by 43%, and Table 6's breakdown of stereo matching (22–36% correct on short videos, 95–100% on long) is a useful data point for anyone trying epipolar matching on fish. The appendices with 3D trajectory visualizations show the intended output. If the paper were reframed as 'we built this pipeline and here is where it fails on short videos,' it would be a decent application note.\n\nSoft spots beyond the central missing comparison. First, the novelty statement in the intro and related work—'none of these models have been used on animal data' and no prior multi-view adaptation—is contradicted by the paper's own citations [2,8,12], which are same-group fish/animal MOT papers. Second, the 'novel 3D output' is standard triangulation. Third, the re-ID thresholds (50 px, 100-frame window, 10-frame overlap) are hand-tuned with no sensitivity analysis. Fourth, code is promised via a GitHub link that does not appear in the text. Fifth, short-video 3D results rely on manual re-matching, so the automatic pipeline's only solid success is on long videos.\n\nThis is a paper for ecologists wanting a starting point for stereo fish tracking, or for anyone looking for a cautionary example of evaluation design. It deserves a serious referee rather than a desk reject, but the authors should be told to either add a real multi-view versus single-view comparison (for example, using 3D information to correct 2D track errors) or drop the accuracy-enhancement claim entirely.","headline":"A workmanlike stereo fish-tracking pipeline whose central accuracy claim is never tested; the real contribution is the dataset and honest evaluation, not the multi-view enhancement.","tokens_in":16863,"tokens_out":3406,"would_cite":false,"duration_ms":23698,"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 stereo pair of off-the-shelf single-view trackers can yield 3D fish trajectories for ecological studies.","keywords":["multi-object tracking","fish tracking","multi-view fusion","stereo matching","YOLOv8","ByteTrack","3D trajectory reconstruction","underwater video analysis"],"falsifier":"Run the pipeline on the short stereo videos without manual re-matching and compare the automatically matched fish-ID pairs to a hand-labelled ground-truth correspondence; if the automatic match rate stays at the reported 22–36% rather than near 100%, the framework's central identity-preservation premise fails for short recordings. A complementary check is to compare HOTA/MOTA on the fused multi-view output against the single-view output on the same videos: the claimed enhanced tracking accuracy would be falsified if the fused trajectories do not score higher on association metrics.","tokens_in":15878,"feed_emoji":"🐟","tokens_out":7268,"duration_ms":46850,"temperature":0.7,"pith_summary":"This paper argues that a multi-view multiple-object-tracking framework can be built by fusing the outputs of ordinary single-view trackers, and that this fusion improves fish tracking in underwater footage while adding a third dimension. The authors run YOLOv8 with ByteTrack on each of two calibrated stereo videos, re-identify fish that the tracker splits into multiple IDs, match the fish across views using epipolar geometry, and triangulate 3D coordinates per fish per frame. The payoff, if the claim holds, is that ecologists can obtain depth-resolved trajectories, speeds, and grouping patterns from stereo cameras using off-the-shelf trackers, without training a dedicated multi-view model. The paper's own results show the pipeline works well on long videos (95–100% stereo-match accuracy) and poorly on short ones (22–36%), where matches had to be corrected by hand.","feed_headline":"Stereo pair of off-the-shelf trackers yields 3D fish paths","feed_subtitle":"A YOLOv8-ByteTrack pipeline on calibrated stereo video produces depth-resolved fish trajectories for ecological studies.","key_machinery":"The load-bearing mechanism is the post-track fusion pipeline that converts two independent 2D tracking outputs into a 3D trajectory set. Each view is processed by YOLOv8 detection with ByteTrack association; a hand-tuned post-track re-identification pass merges fragmented identities using a 50-pixel proximity threshold, a 100-frame disappearance window, and a 10-frame co-occurrence rule; then stereo correspondence is established through the fundamental matrix and epipolar-line distance thresholds, with the most common per-fish match selected across frames; finally, the calibrated camera parameters feed a triangulation step that produces x, y, z coordinates. The epipolar constraint—that a point in one camera lies on a specific line in the other—is the geometric carrier of the cross-view identity matching.","core_discovery":"The central claim is that single-view multi-object trackers can be dual-leveraged—run independently on each stereo view and then fused through geometry—to produce a new 3D output and enhanced tracking accuracy for fish in underwater videos. The framework detects fish with a relative accuracy of about 47% within the ground-truth margin, tracks a subset of fish for full video lengths, and uses the fundamental matrix and epipolar lines to find corresponding fish IDs between views; the most frequent match per fish across frames is chosen, and camera calibration parameters are used to triangulate x, y, z coordinates. In the paper's account, the 3D output enables more realistic speed, depth, and grouping analyses than 2D single-view tracks, and for a select set of videos the multi-view framework overcomes the challenges of small, visually similar fish with rapidly changing appearances. The authors conclude that this adapts off-the-shelf single-view trackers to ecological studies, a gap they identify in prior work.","pith_inferences":["A direct consequence the authors do not spell out: if off-the-shelf single-view trackers plus epipolar fusion suffice for 3D fish tracking, then the expensive joint multi-view training used in models like DETR3D is unnecessary for this ecological scenario; the bottleneck becomes the per-view detector's recall, not the fusion strategy.","The stereo-matching failure on short videos suggests the frequency-based 'most common match' selection needs temporal accumulation: instead of picking the modal match per fish from a 258-frame video, one could accumulate correspondences across the whole recording, or enforce one-to-one matching globally, which might recover the short-video cases that currently require manual re-matching.","The claimed 'enhanced tracking accuracy' is asserted but never measured against the single-view baseline with the same MOT metrics; a reader could test it by computing HOTA/MOTA on the fused tracks versus the per-view tracks, which would isolate whether the 3D step improves association or merely adds coordinates."],"forward_implications":["Ecologists can generate 3D fish trajectories (x, y, z per frame) from calibrated stereo cameras using off-the-shelf single-view trackers and the described fusion steps.","The framework produces an enriched behavioral dataset: the paper reports speed over time, acceleration, path length, spatial distribution, density maps, temporal patterns, and depth over time from the triangulated coordinates.","On long stereo videos (e.g., 3,117-frame recordings 129 and 406) the stereo-matching step reached 95–100% accuracy, indicating the approach is usable without manual correction on extended footage.","Even with a detection recall near 48%, the post-track re-identification step reduced the number of distinct fish IDs by 43.3% on average, consolidating fragmented tracks before fusion."],"supporting_citations":[{"why":"Supplies the YOLOv8 detector used in each single-view branch of the framework.","marker":"[15]"},{"why":"Supplies the ByteTrack association algorithm that links YOLOv8 detections into tracks per view.","marker":"[26]"},{"why":"Supplies the epipolar geometry and fundamental-matrix machinery used for cross-view stereo matching.","marker":"[10]"},{"why":"Defines the HOTA evaluation metric used to assess tracking accuracy, including the reported AssA/DetA decomposition.","marker":"[14]"},{"why":"Defines the MOT ground-truth format and benchmark conventions the fish dataset follows.","marker":"[5]"},{"why":"Serves as the multi-view inspiration for the framework's design, though not directly integrated.","marker":"[21]"}],"fun_headline_variants":["Two views, one track: off-the-shelf MOT for 3D fish","Fusing single-view trackers to track fish in 3D","Stereo MOT via existing single-view models","3D fish paths from stereo and standard MOT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole framework stands on the assumption that the stereo-matching step pairs the same physical fish in the two camera views, so that every triangulated 3D point is the same real fish.","fun_headline_variants_meta":{"raw":{"variants":["Two views, one track: off-the-shelf MOT for 3D fish","Fusing single-view trackers to track fish in 3D","Stereo MOT via existing single-view models","3D fish paths from stereo and standard MOT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001426,"raw_usage":{"total_tokens":5734,"prompt_tokens":909,"completion_tokens":4825,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":4756}},"tokens_in":525,"tokens_out":4825,"duration_ms":29886,"temperature":1.0,"reasoning_tokens":4756,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:50:14.860544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on the short stereo videos without manual re-matching and compare the automatically matched fish-ID pairs to a hand-labelled ground-truth correspondence; if the automatic match rate stays at the reported 22–36% rather than near 100%, the framework's central identity-preservation premise fails for short recordings. A complementary check is to compare HOTA/MOTA on the fused multi-view output against the single-view output on the same videos: the claimed enhanced tracking accuracy would be falsified if the fused trajectories do not score higher on association metrics.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the YOLOv8 detector used in each single-view branch of the framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the multi-view inspiration for the framework's design, though not directly integrated."}],"review_version":1}