{"id":"d0f31da6-d2d8-48e3-9302-c7680ebaf949","arxiv_id":"2505.12130","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A unified bottom-up model that refines keypoints with disk-relative centroid vectors and uses high-confidence keypoints as dynamic mask centers outperforms prior pose and segmentation models on COCO, CrowdPose, and OCHuman.","lead":"KDC is a bottom-up network that estimates human keypoints and then uses those keypoints as dynamic centers to group pixels into instance masks for each person. The authors report state-of-the-art accuracy on COCO, CrowdPose, and OCHuman for pose and segmentation, along with real-time runtime.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"KDC's SOTA numbers are not yet attributable to its proposed components: Table 7 covers only keypoints, Fig. 8 gives no DMc-vs-SMc numbers, and Eq. 5's offset loss is written as an identity. A same-backbone ablation is required to support the central claim.","rationale":"I read the paper's strongest claim as the end-to-end SOTA plus the assertion that the two new mechanisms cause it. The reader's weakest assumption (single-centroid unimodality for MaskCentroid) is plausible but not the most decisive issue; the more immediate problem is that the paper never provides a quantitative isolation of MaskCentroid on the segmentation task and the stated offset loss is a tautology. Those two issues directly undermine the causal attribution in Table 7. I do not claim fraud or that the architecture cannot work; the equations may have typographical errors and missing details. But the manuscript, as written, does not allow a reader to verify that dynamic centroids are learned or that they drive segmentation performance. The proposed check is a standard controlled ablation plus an equation correction; it would settle the concern without requiring new benchmark design. Because the reader's CONDITIONAL verdict already reflects verification weakness, my read does not move the verdict; I would keep it CONDITIONAL pending code/weights and the corrected ablation. I partially agree with the reader: we both doubt MaskCentroid, but I focus on causal attribution and the vacuous loss rather than on the unimodal-cloud assumption.","tokens_in":14559,"tokens_out":6797,"duration_ms":75847,"concrete_test":"On COCO val, train the full RN-152 KDC with Eq. 5 corrected to compare a predicted embedding with the target mi+vi, and train the same model with MaskCentroid fixed to the static instance centroid of Eq. 6 (SMc), all other hyperparameters identical. Report keypoint and person-mask AP for KDC, KDC-SMc, and a no-KeyCentroid baseline. If DMc minus SMc is below 1 AP, or if either run reproduces the reported 47.6-level mask AP only after changing the stated loss, the central attribution fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that KeyCentroid and dynamic MaskCentroid produce the reported gains. Two load-bearing conditions are insecure. First, the only tabulated ablation (Table 7) reports keypoint AP on COCO val; it never quantifies segmentation, and the DMc-vs-SMc comparison is a qualitative figure (Fig. 8) with no AP. Thus the +3.1 mask AP over Pose+Seg on COCO test (Table 4) cannot be traced to the MaskCentroid mechanism; it could come from the ResNet-152 backbone, training budget, or evaluation details. Second, the MaskCentroid training objective is internally inconsistent as written. In §3.3, the paper defines ei = mi + vi and then minimizes L_offset = (1/N) Σ || ei - (mi + vi) ||_1, which is identically zero because the two terms are equal by definition. No gradient for the offset branch follows from this loss. The subsequent Eq. 7 replaces the centroid with the mean of the assigned pixel embeddings, but assignment depends on the very embedding whose training signal is undefined. A reader cannot tell whether the network actually learned dynamic clustering or whether the segmentation head is trained by an unstated loss. These gaps are load-bearing because the paper's novelty and SOTA margins rest on exactly these components.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KDC, a bottom-up unified model for human pose estimation and instance-level segmentation. PoseNet detects keypoints via a disk-based heatmap representation and refines them with a KeyCentroid mechanism, while SegNet uses high-confidence keypoints as dynamic MaskCentroids to cluster pixels in an embedding space. Experiments on COCO, CrowdPose, and OCHuman report strong or state-of-the-art accuracy, with real-time speed and no person detector. The central claim is that the KeyCentroid and dynamic MaskCentroid mechanisms are responsible for the reported gains.","tokens_in":14897,"tokens_out":4631,"duration_ms":47167,"significance":"If the reported results hold, KDC would be a valuable real-time, bottom-up unified pose-and-segmentation model; the keypoint results on COCO and CrowdPose are competitive, and the runtime/parameter comparisons are useful. The paper provides a quantitative ablation for the keypoint components (Table 7). However, the evidence for the MaskCentroid contribution is only qualitative, and the training objective for the offset branch is written in a degenerate form. The significance of the segmentation claims therefore cannot be assessed from the manuscript as it stands.","major_comments":[{"comment":"The offset loss is identically zero as written: e_i is defined by e_i = m_i + v_i, so ||e_i - (m_i + v_i)||_1 is always 0. This means the offset branch receives no gradient and the actual training signal for the pixel embeddings is unspecified. The authors must state the real loss and clarify how e_i is generated by the network.","section":"Section 3.3, Eq. (5)"},{"comment":"The only quantitative ablation (Table 7) reports keypoint AP; the SMc-vs-DMc comparison is presented only as a qualitative figure (Fig. 8) with no segmentation AP. Consequently, the +3.1 mask AP over Pose+Seg in Table 4 cannot be attributed to the MaskCentroid mechanism. A same-backbone, same-training ablation reporting mask metrics is needed to support the central claim.","section":"Section 5.2 and Table 7"},{"comment":"Table 6 compares KDC's 2D pose AP50 to projected 3D pose models (CRMH, ROMP) on CrowdPose. This conflates different tasks and evaluation protocols; the large margin is not a meaningful comparison. The authors should remove this comparison or explicitly label it as a different-task illustration rather than a competitive benchmark.","section":"Table 6"},{"comment":"The 'dynamic movement' motivation is never measured: all benchmarks are static images, and the dynamic-centroid benefit under rapid motion is only inferred from single-frame results. Please either add temporal/video experiments or qualify the claim that dynamic centroids help during rapid body movements.","section":"Section 1 and Section 3.3"},{"comment":"Equation (7) defines the dynamic centroid as the mean of assigned pixel embeddings, but assignment depends on distances to that same centroid. The paper does not analyze this self-referential clustering, nor does it test the single-centroid assumption for highly occluded or disconnected person instances, which is directly relevant to the OCHuman evaluation. Add an experiment or analysis that isolates this assumption.","section":"Section 3.3, Eqs. (6)-(7)"}],"minor_comments":[{"comment":"The contribution list contains a typo: 'developement' should be 'development'.","section":"Section 1"},{"comment":"The learning rate is written as '0.1×e-4', which is ambiguous; it should be written as 1e-5 or similar.","section":"Section 4"},{"comment":"Several references are malformed, e.g., [George et al., 2018] and [Golda and others, 2019]; the author lists and citation keys should be corrected.","section":"References"},{"comment":"Figures 8-10 are screenshots without clear axis labels or quantitative legends; they should be redrawn as standard plots so the ablation results are legible.","section":"Figures 8-10"}],"recommendation":"major_revision","confidential_remarks":"The identity-loss issue in Eq. (5) is severe enough that the method section needs a rewrite, and the missing segmentation ablation is load-bearing. I recommend major revision rather than rejection because the keypoint ablation (Table 7) is sound and the problems are localizable. I also note that the closest baselines, Pose+Seg and VisualCent, are from the authors' own group; the comparisons would benefit from independent replication or at least a clear statement of shared training details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2505.12130. The interesting part is real but the verification is not. KDC recombines two known ideas — disk-based offset regression (PersonLab-style) and nearest-centroid embedding clustering (associative embeddings) — and anchors mask centroids at high-confidence keypoints. That's a sensible design choice and the benchmark numbers are eye-catching: 76.1 keypoint AP and 47.6 mask AP on COCO with ResNet-152, plus strong CrowdPose/OCHuman results. If those numbers reproduce, it's a useful contribution to bottom-up unified human analysis.\n\nWhat the paper does well: the pipeline is clearly described, the three-benchmark evaluation is a reasonable scope, and the runtime comparison in Fig. 6 gives a concrete sense of efficiency. The idea of using keypoint detections as cluster centers for segmentation is worth exploring.\n\nNow the soft spots, in order of severity.\n\nFirst, the MaskCentroid training loss is broken as written. In §3.3, e_i is defined as m_i + v_i, so L_offset = (1/N) Σ ||e_i − (m_i + v_i)||_1 is identically zero. That cannot train the offset branch. Either there is a typo (the network output should be e_i and m_i + v_i the target) or the segmentation head has no defined training signal. Since MaskCentroid is the paper's main novelty, this is load-bearing.\n\nSecond, the ablation doesn't support the attribution. Table 7 only reports keypoint AP; there is no segmented AP table for static vs. dynamic MaskCentroid. Fig. 8 is a qualitative comparison. So the +3.1 mask AP over Pose+Seg in Table 4 can't be traced to the dynamic centroid mechanism; it could come from the ResNet-152 backbone, training budget, or other details. A same-backbone ablation with numbers is needed.\n\nThird, Table 6 compares a 2D pose model to projected 3D models under AP50. That's not a fair baseline and should be removed or reframed.\n\nMinor: the per-keypoint Gaussian sigmas are hand-tuned to the benchmarks, no code or error bars, and the 'rapid dynamic movement' motivation is never tested — all evaluation is on static images.\n\nOverall: the engineering idea has merit and the reported numbers are worth checking, but the central loss is formally undefined and the empirical claims aren't yet supported by the ablation. I would engage with this as a reviewer: send it out, ask for a corrected formulation, a proper DMc-vs-SMc ablation, and code. It's not desk-reject material, but it's not ready in this form.","headline":"A plausible bottom-up pose+segmentation idea, but the MaskCentroid loss is an identity as written and the ablation doesn't support the key claims.","tokens_in":15361,"tokens_out":3770,"would_cite":false,"duration_ms":35855,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that using detected keypoints as dynamic cluster centers lets a single bottom-up network produce accurate human joint locations and instance-level body segmentation in real time, without a person detector.","keywords":["human pose estimation","instance segmentation","bottom-up","dynamic centroid","KeyCentroid","MaskCentroid","crowd scenes","occlusion"],"falsifier":"Run KDC on video frames where a person's visible body parts are separated by occlusion or a barrier, and track whether the dynamic centroid follows a fast-moving joint. If one person's pixels split into two clusters, or the mask lags when a joint moves quickly, the single-centroid assumption fails; the paper reports no such dynamic or disjoint-occlusion test.","tokens_in":14377,"feed_emoji":"🧍","tokens_out":12091,"duration_ms":104415,"temperature":0.7,"pith_summary":"The paper tries to establish that one bottom-up network can locate a person's joints and cut out the person's silhouette at the same time, without first running a separate person detector. The proposed KDC model treats detected keypoints as dynamic centroids: KeyCentroid refines joint coordinates inside a local disk, then MaskCentroid uses the high-confidence joints as moving cluster centers that pull mask pixels to the correct person in an embedding space. On COCO it reports a keypoint mAP of 76.1 and a segmentation mAP of 47.6 with a ResNet-152 backbone, along with strong results on the crowded CrowdPose and heavily occluded OCHuman benchmarks, while running in real time. If the claim holds, unified pose-and-segmentation becomes practical for live and crowded scenes without a detector's cost.","feed_headline":"Joints as cluster centers: 76.1 pose, 47.6 mask on COCO","feed_subtitle":"The same bottom-up network produces joints and masks in real time, no detector needed.","key_machinery":"The load-bearing machinery is the centroid pair. KeyCentroid places a disk $D_R(q)=\\{p:\\|p-q\\|\\le R\\}$ around each joint; every pixel inside regresses a vector toward the joint's centroid, and pointwise Gaussian smoothing turns these into the final keypoint coordinates and confidence scores. MaskCentroid works in a learned embedding space where each mask pixel stores an offset toward its person's centroid; high-confidence keypoints act as dynamic centroids $C_i$, and pixel-to-instance assignment uses the Gaussian $\\phi_j(e_i)=\\exp(-\\|e_i-C_i\\|^2/2\\sigma_j^2)$, with the dynamic variant (Eq. 7) replacing the centroid by the mean embedding. The combination lets one network cluster pixels to instances without a person detector, while the cluster centers can shift as keypoints move.","core_discovery":"The paper's central claim is that a common centroid mechanism solves both tasks better than prior joint models. PoseNet produces keypoint heatmaps in a disk representation and uses KeyCentroid to regress each pixel inside the disk toward the joint's center, which the ablations credit for improved localization and confidence scores. SegNet then takes the resulting high-confidence keypoints as MaskCentroids, dynamic centers of attraction in the embedding space, and assigns each mask pixel to the instance whose centroid it is nearest. The reported gains—COCO keypoint AP 76.1 and segmentation AP 47.6, CrowdPose 74.5, OCHuman 46.3/46.0 keypoint and 58.3/59.6 segmentation—are attributed specifically to these two mechanisms rather than to the backbone, as the ablation study in Table 7 and the static-versus-dynamic comparison in Section 5.2 argue.","pith_inferences":["The centroid mechanism is task-agnostic in principle: any articulated object with annotated keypoints and masks (animals, hands, robots) could use the same keypoint-driven clustering.","The paper motivates 'rapid dynamic movement' but evaluates only on static images; a video experiment with fast articulation would directly test whether the dynamic centroid actually tracks motion better than a static one.","The single-centroid-per-person assumption (Eq. 6) could be relaxed to a mixture of centroids per instance, which may better handle a person split into disconnected visible body parts, a case not covered by the benchmarks.","The disk radius R is fixed (R=32) in all reported experiments; making R depend on image scale or keypoint type could improve small- and large-person variation beyond the sigma tuning shown in Section 5.3."],"forward_implications":["A bottom-up, detector-free model reports higher COCO keypoint and segmentation AP than the top-down Mask R-CNN baseline, suggesting a person detector is not necessary for accurate joint pose plus mask output.","Because keypoints double as clustering centers, the network produces joints and masks from the same features, simplifying the architecture compared to separate pose and segmentation branches.","The reported runtime places the unified model in real-time territory with fewer parameters than Mask R-CNN, PersonLab, and Pose+Seg (Figure 6), opening live-use applications.","On the heavily occluded OCHuman benchmark the reported keypoint mAP of 46.3 (val) and 46.0 (test) exceeds the listed bottom-up baselines, indicating the grouping survives high occlusion levels."],"supporting_citations":[{"why":"The PosePlusSeg predecessor that KDC extends; its static-centroid segmentation is the direct baseline KDC reports beating on both COCO keypoints and segmentation.","marker":"[Ahmad et al., 2022]"},{"why":"PersonLab, the closest bottom-up joint pose-and-segmentation model, whose single- and multi-scale COCO results anchor KDC's keypoint and segmentation comparisons.","marker":"[George et al., 2018]"},{"why":"Mask R-CNN, the standard top-down joint model whose COCO segmentation AP and compute profile serve as a primary baseline in Tables 1, 4, and Figure 6.","marker":"[He et al., 2017]"},{"why":"Supplies the CrowdPose benchmark used to evaluate crowded-scene keypoint accuracy.","marker":"[Li et al., 2019]"},{"why":"Supplies the COCO dataset used for end-to-end training and the main keypoint and instance-segmentation test sets.","marker":"[Lin et al., 2014]"},{"why":"Provides the OCHuman heavily-occluded benchmark and the Pose2Seg detection-free segmentation baseline that KDC reports surpassing by 3.9/4.4 AP.","marker":"[Zhang et al., 2019]"}],"fun_headline_variants":["Keypoints as moving centroids unify pose and segmentation","Dynamic centroids turn joints into mask seeds live","One bottom-up net: pose and masks via joint centroids","Centroid-driven model hits 76.1 pose, 47.6 mask on COCO","No detector: keypoint centroids anchor real-time masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes every person can be captured as one coherent blob of pixels clustered around a single center, even when the person is heavily occluded or moving fast, and the paper never directly tests that assumption in motion or with split-up bodies.","fun_headline_variants_meta":{"raw":{"variants":["Keypoints as moving centroids unify pose and segmentation","Dynamic centroids turn joints into mask seeds live","One bottom-up net: pose and masks via joint centroids","Centroid-driven model hits 76.1 pose, 47.6 mask on COCO","No detector: keypoint centroids anchor real-time masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1363,"prompt_tokens":926,"completion_tokens":437,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":349}},"tokens_in":542,"tokens_out":437,"duration_ms":4838,"temperature":1.0,"reasoning_tokens":349,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:40:10.678527+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run KDC on video frames where a person's visible body parts are separated by occlusion or a barrier, and track whether the dynamic centroid follows a fast-moving joint. If one person's pixels split into two clusters, or the mask lags when a joint moves quickly, the single-centroid assumption fails; the paper reports no such dynamic or disjoint-occlusion test.","supporting_citations":[{"cited_title":"Joint Human Pose Estimation and Instance Segmentation with PosePlusSeg","cited_arxiv_id":null,"evidence_quote":"The PosePlusSeg predecessor that KDC extends; its static-centroid segmentation is the direct baseline KDC reports beating on both COCO keypoints and segmentation."},{"cited_title":"Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model","cited_arxiv_id":null,"evidence_quote":"PersonLab, the closest bottom-up joint pose-and-segmentation model, whose single- and multi-scale COCO results anchor KDC's keypoint and segmentation comparisons."},{"cited_title":"Mi- crosoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"Supplies the COCO dataset used for end-to-end training and the main keypoint and instance-segmentation test sets."},{"cited_title":"Pose2seg: Detection free human instance segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the OCHuman heavily-occluded benchmark and the Pose2Seg detection-free segmentation baseline that KDC reports surpassing by 3.9/4.4 AP."}],"review_version":1}