{"id":"c5cfc062-32d7-49cd-bdcb-8bf16a0359ff","arxiv_id":"2501.11288","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PD-SORT achieves higher HOTA than its OC-SORT baseline on DanceTrack, MOT17, and MOT20 by adding pseudo-depth states to the Kalman filter and using depth-volume IoU and quantized depth costs in data association.","lead":"PD-SORT adds a pseudo-depth cue to the standard multi-object tracking pipeline by extending the Kalman filter state with depth-like coordinates and using those cues in matching. This improves identity consistency under occlusion on DanceTrack and, to a lesser extent, on MOT17 and MOT20, at a modest speed cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-depth is an affine transform of the box-bottom row; QPDM is invariant under that transform, so the 'depth' state supplies no information beyond the 2D box and the KF treats a dependent quantity as independent.","rationale":"The reader's weakest assumption correctly identifies the flat-ground premise and the unmodeled dependency in the Kalman filter. My concern is more fundamental: even under an ideal flat ground, pd is a deterministic affine function of Yb, so the pseudo-depth state adds no independent information. The QPDM cost is invariant to that affine transform, making it literally a quantized bottom-row consistency cost. This weakens the paper's interpretive claim that depth provides new information, and it makes the KF extension internally inconsistent because a derived coordinate is modeled as an independent state. However, the empirical claim is not falsified: the reported test-set numbers come from official benchmark evaluation, and the ablations are directionally consistent with the modules helping. The algorithmic typos in Eq. 4 and Algorithm 1 (missing parentheses, max(x1_1 - x2_1), and minprevious never updated) further justify a conditional verdict but are secondary to the redundancy issue. The appropriate response is to keep the CONDITIONAL verdict and require a control experiment using Yb directly, clearer derivation of the KF measurement model, and runnable code that matches the published equations.","tokens_in":18714,"tokens_out":9492,"duration_ms":105266,"concrete_test":"Run the QPDM ablation from Table IV (baseline + QPDM) with the pseudo-depth value replaced by -Yb, so the same min-max normalization and interval assignment are applied to the raw bottom coordinate; if the HOTA/AssA deltas match the published pd-based numbers, then QPDM is exactly quantized bottom-row consistency and the depth story is a relabeling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Pseudo-depth is defined in Eq. 1 as pd = 2*IMGh - Yb, so it is an affine, monotonically decreasing function of the detection's bottom row Yb. Yb is part of the 2D bounding box already used by every SORT-family tracker; hence PD-SORT adds no new measurement to the observation stream. The KF state in Eq. 3 appends pd and vpd as if they were independent states, but the observation pd is exactly determined by the center y-coordinate and box height (h = sqrt(s/r)). The filter's diagonal covariance therefore encodes a constraint that the data-generating process does not have, making the augmented motion model internally inconsistent. In QPDM, depths are min-max normalized frame-wise (Eq. 7), and min-max normalization is invariant to affine monotone transformations; replacing pd with Yb yields exactly the same quantized interval costs up to the order of intervals. QPDM is therefore indistinguishable from quantized bottom-row consistency. DVIoU is a scalar reweighting of the same 2D boxes by a function of Yb, not a 3D volume computed from independent depth evidence. Even when the flat-ground assumption is perfectly satisfied, the 'depth' state carries no information beyond the 2D box. The benchmark numbers may be real, but they do not demonstrate that pseudo-depth supplies new discriminative information; the occlusion-robustness mechanism attributed to depth is at minimum mislabeled.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PD-SORT, an online, real-time, motion-based multi-object tracker built on OC-SORT. The authors extend the Kalman filter state with a pseudo-depth value defined as an affine function of the bounding box bottom row, introduce a Depth Volume IoU (DVIoU) that reweights 2D IoU by this pseudo-depth, and add a Quantized Pseudo-Depth Measurement (QPDM) cost based on min-max normalized, interval-quantized pseudo-depth values. Camera motion compensation is also integrated. The method is evaluated on DanceTrack, MOT17, and MOT20 under the private-detection protocol, reporting consistent improvements over OC-SORT, most notably +3.6 HOTA on DanceTrack, alongside component ablations on the DanceTrack validation set.","tokens_in":18910,"tokens_out":3761,"duration_ms":37666,"significance":"If the central claim is taken at face value, the paper would provide a simple, real-time, appearance-free tracker that improves occlusion robustness by integrating a depth-like cue into the motion model. The empirical work is substantial: benchmark results are from official MOT servers, the baseline and proposed method share the same detections, and the ablations show monotone gains from each added module. The paper also ships code and reports runtime. However, the significance is substantially qualified by the fact that the pseudo-depth is a deterministic affine transform of the bounding box bottom row, which is already implied by the existing Kalman state. This makes the 'depth' interpretation, and the claim that depth supplies new discriminative information, not supported by the current presentation. The observed gains may be real and reproducible, but they are better described as gains from an engineered bottom-row consistency feature rather than from a genuinely independent depth measurement.","major_comments":[{"comment":"The pseudo-depth defined as pd = 2*IMGh - Yb is an affine, monotonically decreasing function of the detection's bottom row Yb. Since Yb is fully determined by the existing Kalman state (xc, yc, s, r) through the box relation (Yb = yc + 0.5*sqrt(s/r)), the augmented state X = [xc, yc, pd, s, r, vx, vy, vpd, vs] treats a derived coordinate as an independent state. The Kalman filter's diagonal process and measurement noise covariances therefore encode a constraint that does not match the data-generating process; the augmented state is internally redundant. The paper should either provide a principled justification for why the derived coordinate should be modeled as an independent state with its own velocity, or reframe the contribution as using bottom-row consistency within the existing state space.","section":"Section III-A, Eq. (1) and Eq. (3)"},{"comment":"The min-max normalization in QPDM is invariant to affine monotone transformations of the input. Because pd = C - Yb with constant C = 2*IMGh, normalizing pd yields (Yb_max - Yb)/(Yb_max - Yb_min), which is exactly the reversed normalized bottom row. Consequently, QPDM is indistinguishable from quantized bottom-row consistency; it does not use depth information beyond what is already present in the 2D box. The ablation in Table V demonstrates that quantization helps, but it does not support the claim that the pseudo-depth interpretation provides extra discriminative information. The authors should add a comparison where the same QPDM pipeline is applied directly to Yb, or to a genuinely independent depth estimate, to isolate the source of the gains.","section":"Section III-C, Eq. (7) and Algorithm 1"},{"comment":"DVIoU is not a three-dimensional volume IoU in any geometric sense; the 'depth volume' is the 2D intersection area multiplied by a scalar that is an affine function of the box bottom row (pd = C - Yb). Thus DVIoU is a reweighting of the standard 2D IoU by a Yb-dependent factor, not an intersection over union computed in a space with an independent depth axis. The improvement over standard IoU in Table VI may be a legitimate empirical effect, but the 'depth volume' terminology is misleading. The authors should either rename DVIoU to reflect its actual form or provide evidence that the geometric interpretation as a 3D volume is consequential, for example by comparing against an equivalent reweighting based directly on Yb.","section":"Section III-B, Eq. (4)-(6)"},{"comment":"The central claim that pseudo-depth 'significantly alleviates the occlusion-induced ambiguous associations' implies that depth is a new cue. Given that the pseudo-depth is a deterministic transform of an existing 2D box coordinate, the benchmark improvements on DanceTrack do not by themselves demonstrate that a depth cue is responsible. A concrete test would be to rerun PD-SORT with the pseudo-depth everywhere replaced by Yb (or by an independent monocular depth estimate) and compare the metrics. Without such an experiment, the novelty claim that this is the first work to incorporate depth as a state in a pure motion-based 2D MOT is overstated.","section":"Abstract and Section I"}],"minor_comments":[{"comment":"The phrase 'We inherent the observation-centric idea' should be 'We inherit'; the same typo appears in Section III-E.","section":"Section I, last paragraph of Introduction"},{"comment":"The formula for winter contains a typo: 'max(x1_1 - x2_1)' should be 'max(x1_1, x2_1)' (the max of the two left coordinates), otherwise the expression is dimensionally inconsistent.","section":"Eq. (4)"},{"comment":"In the QPDM pseudocode, 'minprevious ← 1' inside the loop should be 'minprevious ← mincurrent' (or the equivalent) to advance the interval lower bound; as written, all intervals are tested against [1, mincurrent] repeatedly, which does not match the interval subdivision described in the text.","section":"Algorithm 1, lines 7 and 15"},{"comment":"The text reads 'the pesudo-depth' – spelling should be 'pseudo-depth'.","section":"Section IV-C-1, first paragraph"},{"comment":"The sentence 'we test the frames per second (FPS) of our method (28.7 FPS) and the baseline (35.1 FPS) on on the same device' contains a duplicated 'on' and should be reworded.","section":"Section IV-B-2, computational efficiency"},{"comment":"The abbreviation for Depth Volume IoU is inconsistently written as 'DvIoU' in the section heading and 'DVIoU' elsewhere; the notation should be unified.","section":"Section III-C, heading and text"}],"recommendation":"major_revision","confidential_remarks":"The paper reports credible official benchmark numbers and a clean ablation chain, and the code release is a plus. The central weakness is not the experiments but the interpretation: pseudo-depth as defined is a deterministic function of the existing 2D state, so the paper's claim to introduce depth as a new motion state is not supported by the evidence as presented. A revision that honestly repositions the contribution as a bottom-row consistency feature, or that adds a direct comparison against Yb-based features to justify the depth framing, would make the paper publishable. I would not reject outright because the empirical gains appear real and the engineering could be of interest to the tracking community, but the current framing overstates the novelty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the benchmark numbers are real and the ablation logic is sound, but the central \"depth\" claim is overstated. pd = 2H - yb, so it's an affine copy of the bottom row; QPDM min-max normalizes, which cancels affine transforms, so the quantized depth cost is literally quantized bottom-row consistency. The stress-test note is right on the algebra. That said, this doesn't make the paper useless. The specific cost formulations—DVIoU and QPDM—still change the matching behavior in ways that help on DanceTrack, and the gains come from official evaluation servers. The ablations show each module contributes, and the code seems to be available.\n\nWhere it's soft: (1) The KF state augmentation treats pd and vpd as independent when they're deterministically determined by the box; that's not just a framing issue, it's an inconsistent noise model, though in practice it may behave like a regularizer. (2) The paper claims to extend 2D MOT to 3D, but there's no new measurement entering the system; the pseudo-depth is computed from the same detection box. Calling it a depth volume IoU is generous. (3) Algorithm 1 has obvious errors—line 7 resets minprevious to 1 instead of mincurrent, and the normalization lacks parentheses—which would trip up a reimplementation. Eq. 4 has a typo in the width term. (4) The hyperparameters (λ1, λ2, IoU thresholds) are tuned per dataset, and ablations are only on DanceTrack, so the generalization claim rests on the test-set numbers alone.\n\nI don't think this is a fatal problem. The method could be reframed as using the bottom row as a complementary matching cue, and the empirical results would likely stand. The paper is honest about limitations and doesn't hide the small MOTA drop on MOT20. For an editor: this deserves peer review; the errors are fixable and the idea, though less novel than claimed, is a useful addition to the SORT-lineage literature. The right outcome is probably a conditional accept after a careful revision reframes the contribution.","headline":"Solid incremental MOT work with honest benchmark gains, but the pseudo-depth is a transform of the box bottom and the paper oversells the novelty of the depth cue.","tokens_in":19545,"tokens_out":3342,"would_cite":false,"duration_ms":34004,"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":"Pseudo-depth cues added to a Kalman filter resolve occlusion ambiguities in multi-object tracking, lifting DanceTrack HOTA from 54.6 to 58.2.","keywords":["multi-object tracking","tracking-by-detection","pseudo-depth","Kalman filter","data association","occlusion robustness","depth volume IoU","quantized pseudo-depth measurement"],"falsifier":"Run PD-SORT against its baseline on a scene where the camera is at ground level looking up at objects on a balcony or stairs, so that box-bottom position no longer tracks true distance; if HOTA and AssA gains disappear or reverse, the reported improvement is coming from the flat-ground pseudo-depth cue.","tokens_in":18423,"feed_emoji":"🎯","tokens_out":9187,"duration_ms":84570,"temperature":0.7,"pith_summary":"This paper tries to show that a cheap depth proxy, read off from each 2D bounding box, can resolve the ambiguous associations that break tracking-by-detection under occlusion. PD-SORT adds pseudo-depth and its velocity to the Kalman filter state, replaces ordinary 2D IoU with a depth-volume IoU, and adds a quantized pseudo-depth cost to the matching step. On DanceTrack, where targets look alike, move nonlinearly, and constantly occlude one another, PD-SORT raises HOTA from 54.6 to 58.2 relative to its OC-SORT baseline; gains on MOT17 and MOT20 are smaller but consistent. If the claim holds, occlusion-prone tracking can be improved without depth sensors, appearance models, or slower pipelines, using only geometry already present in the 2D boxes.","feed_headline":"Pseudo-depth from 2D boxes sharpens occlusion-ridden tracking","feed_subtitle":"Adding a depth-like coordinate to the Kalman filter lifts DanceTrack HOTA from 54.6 to 58.2 with no depth sensor.","key_machinery":"The load-bearing mechanism is the complementary-view pseudo-depth, $pd = 2 \\times \\mathrm{IMG}_h - Y_b$, which maps a box's bottom-row coordinate to a distance-like quantity that never goes negative at image boundaries. This single number is reused in three places: it is appended to the Kalman state along with its velocity $v_{pd}$; it becomes the third dimension in DVIoU, turning 2D area overlap into volume overlap; and it is binned into intervals by QPDM, so association compares coarse depth order rather than potentially noisy exact values. The same pseudo-depth also makes the recovery stage more selective, because DVIoU between a lost tracklet's last observation and a candidate detection now requires depth agreement as well as spatial overlap.","core_discovery":"The central claim is that a purely geometric depth proxy, computed from each 2D bounding box, can carry much of the association burden in tracking-by-detection when appearance is uninformative and motion is non-linear. PD-SORT takes the pseudo-depth $pd = 2 \\times \\mathrm{IMG}_h - Y_b$ (distance from the box bottom to the bottom of an imaginary complementary view) and treats it as a real motion state: the Kalman state grows from $[x_c, y_c, s, r, v_x, v_y, v_s]$ to $[x_c, y_c, pd, s, r, v_x, v_y, v_{pd}, v_s]$. It then replaces 2D IoU with depth-volume IoU (DVIoU) in both the regular association and the lost-tracklet recovery, and adds a quantized pseudo-depth measurement (QPDM) that compares which of eight depth intervals each tracklet and detection falls into rather than comparing raw depth values. On DanceTrack test the result is HOTA 58.2 versus 54.6 for the OC-SORT baseline, with more modest gains on MOT17 and MOT20; the paper attributes the margin specifically to the depth cue disambiguating boxes that overlap heavily in 2D but sit at different depths.","pith_inferences":["My inference: because QPDM compares only coarse depth intervals, the active ingredient may be relative depth order rather than absolute depth; the authors do not test this separation.","My inference: the flat-ground assumption means the method's advantage should shrink on drone or elevated-camera footage; a direct benchmark there would define the method's operating envelope.","My inference: replacing the geometric pseudo-depth with a learned monocular depth estimate would test whether depth quality is the ceiling; if gains grow, the paper's cue is a floor, not a limit."],"forward_implications":["Motion-only trackers inherit an occlusion-resolving cue without adding detectors, ReID models, or depth sensors.","Because the changes are confined to the Kalman state and the cost matrix, other SORT-family trackers should be able to adopt the same pseudo-depth extension directly.","The DanceTrack margin (+3.6 HOTA) indicates that in uniform-appearance, nonlinear-motion scenes, much of the association error was resolvable by depth order alone.","The modest MOT17 and MOT20 gains suggest the main benefit appears where 2D overlap is ambiguous, while simpler scenes gain only slightly."],"supporting_citations":[{"why":"It supplies the OC-SORT baseline, including the observation-centric modules OCM, OCR, and ORU that PD-SORT inherits, and every reported comparison is measured against it.","marker":"[10]"},{"why":"It introduces the projection-based pseudo-depth idea from which the paper's complementary-view variant is derived.","marker":"[15]"},{"why":"It defines the SORT Kalman state and the tracking-by-detection pipeline that PD-SORT extends with pseudo-depth states.","marker":"[6]"},{"why":"It provides the camera motion compensation technique that PD-SORT applies to correct Kalman states and historical observations.","marker":"[18]"},{"why":"It is the DanceTrack benchmark whose uniform appearances, nonlinear motions, and frequent occlusions produce the large reported HOTA gain.","marker":"[14]"},{"why":"It is the MOT17 benchmark used to show the method generalizes beyond DanceTrack-like motion.","marker":"[42]"},{"why":"It is the MOT20 crowded-scene benchmark used to show the method generalizes to dense scenes.","marker":"[43]"}],"fun_headline_variants":["PD-SORT: Pseudo-depth cue solves occlusion ambiguity in MOT","No depth sensor needed: pseudo-depth sharpens MOT","Box-based depth proxy boosts tracking scores on DanceTrack","Depth from boxes: PD-SORT lifts HOTA to 58.2 on DanceTrack","Pseudo-depth in Kalman filter tames occlusion chaos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the camera being above a flat ground plane with all tracked objects standing on it, so that the bottom edge of each bounding box honestly encodes distance; if that geometry fails, the pseudo-depth cue becomes misleading.","fun_headline_variants_meta":{"raw":{"variants":["PD-SORT: Pseudo-depth cue solves occlusion ambiguity in MOT","No depth sensor needed: pseudo-depth sharpens MOT","Box-based depth proxy boosts tracking scores on DanceTrack","Depth from boxes: PD-SORT lifts HOTA to 58.2 on DanceTrack","Pseudo-depth in Kalman filter tames occlusion chaos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000477,"raw_usage":{"total_tokens":2426,"prompt_tokens":1071,"completion_tokens":1355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":1267}},"tokens_in":687,"tokens_out":1355,"duration_ms":12955,"temperature":1.0,"reasoning_tokens":1267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:26:38.033180+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PD-SORT against its baseline on a scene where the camera is at ground level looking up at objects on a balcony or stairs, so that box-bottom position no longer tracks true distance; if HOTA and AssA gains disappear or reverse, the reported improvement is coming from the flat-ground pseudo-depth cue.","supporting_citations":[{"cited_title":"Observation- centric sort: Rethinking sort for robust multi-object tracking,","cited_arxiv_id":null,"evidence_quote":"It supplies the OC-SORT baseline, including the observation-centric modules OCM, OCR, and ORU that PD-SORT inherits, and every reported comparison is measured against it."},{"cited_title":"Simple online and realtime tracking,","cited_arxiv_id":null,"evidence_quote":"It defines the SORT Kalman state and the tracking-by-detection pipeline that PD-SORT extends with pseudo-depth states."},{"cited_title":"Dancetrack: Multi-object tracking in uniform appearance and diverse motion,","cited_arxiv_id":null,"evidence_quote":"It is the DanceTrack benchmark whose uniform appearances, nonlinear motions, and frequent occlusions produce the large reported HOTA gain."}],"review_version":1}