{"id":"5c88d905-4a49-4fa9-b8d6-558c59604a39","arxiv_id":"2504.19032","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"VISUALCENT reports new state-of-the-art bottom-up human pose and instance segmentation numbers on COCO and OCHuman using disk-based keypoint heatmaps, per-keypoint offset fields, and dynamic keypoint-anchored mask clustering.","lead":"VISUALCENT is a bottom-up neural network that detects human body keypoints and produces a per-person outline in one pass, using predicted keypoints as moving anchor points for pixel grouping. It reports higher accuracy and faster runtime than several prior pose-and-segmentation systems on COCO and OCHuman benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MaskCentroid clustering rule and its training loss are never specified, so the central segmentation claim cannot be checked from the manuscript.","rationale":"I read the paper in good faith. The headline numbers are internally arithmetically consistent for the ResNet-152 rows, and the proposed bottom-up design is plausible. The most serious problem is not an implausible number but an underspecified mechanism. The MaskCentroid equations are circular without a clustering or assignment algorithm, and no loss function is given for the segmentation head. Since MaskCentroid is the advertised contribution behind the segmentation improvements, the central claim depends on an algorithm that the manuscript does not define. The reader identified a related but downstream risk: occlusion can make the single anchor keypoint unusable. My concern is more basic: even in the non-occluded case, the paper never states how instances are formed or trained. This justifies keeping the verdict conditional, with the condition being a precise specification and reproducible implementation. The proposed check, a static-versus-dynamic anchor ablation using the released code, would settle whether the dynamic MaskCentroid mechanism is actually responsible for the reported gains.","tokens_in":8814,"tokens_out":5605,"duration_ms":65061,"concrete_test":"Obtain the released code and the exact MaskCentroid loss plus inference assignment rule, then run a controlled ablation on COCO val with the dynamic high-confidence keypoint anchor versus a static per-instance centroid, keeping all other settings identical. If the dynamic anchor does not reproduce the Fig. 6 / Table III margin, or if the loss and assignment cannot be instantiated from Eqs. (2)-(3), the central segmentation claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing weak point is not an arithmetic slip but a formal gap in the proposed segmentation mechanism. Section III defines MaskCentroid via Eq. (2) as F' = (1/|N_j|) sum_{m_i in N_j} m_i, and Eq. (3) as a Gaussian similarity to (1/|N_j|) sum_{e_j in N_j} e_j. Both formulas presuppose that the instance N_j is already known; no inference-time assignment rule explains how a predicted high-confidence keypoint anchors pixels. No training loss for the 'MaskCentroid offset' is written; the text only says losses are 'carefully balanced.' The phrase 'defining the highly confident keypoint as a learnable parameter' is not an algorithm. Consequently, the paper's central segmentation claim, that dynamic MaskCentroids improve mask AP over PosePlusSeg and PersonLab, cannot be reproduced or even falsified from the manuscript. The reader's anchor-failure scenario is a downstream symptom of this gap: if the missing assignment rule relies on a single keypoint, occluded or missing keypoints have no fallback, but the more basic problem is that the manuscript never supplies the rule to test.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VISUALCENT, a bottom-up framework for unified multi-person 2D pose estimation and instance segmentation. The method first detects keypoints using a keypoint heatmap with a disk representation and a 'KeyCentroid' regressed offset field, then groups mask pixels to instances using a 'MaskCentroid' defined as a high-confidence keypoint that serves as a dynamic clustering anchor. Evaluations are reported on COCO keypoint test-dev, COCO segmentation test, and OCHuman, with tables showing improvements over prior bottom-up methods such as PosePlusSeg and PersonLab. The abstract also claims real-time performance. The central technical novelty is the dynamic MaskCentroid, which is claimed to outperform a static centroid under occlusion and rapid motion.","tokens_in":9079,"tokens_out":3355,"duration_ms":34662,"significance":"If the reported results are reproducible and the dynamic-centroid mechanism is fully specified, the paper would make a useful contribution: it is a detector-free unified pose-and-segmentation model that reports competitive mAP on COCO and particularly strong gains on OCHuman, a highly occluded benchmark. The external evaluation on standard benchmarks and the direct comparisons with recent bottom-up methods are strengths. However, the paper's central claim about the dynamic MaskCentroid is currently under-specified, and the key ablation supporting it contains no numbers. The real-time claim is also not quantified. These gaps prevent the reader from verifying or building on the method, so the current version does not meet the evidentiary bar for its headline claims.","major_comments":[{"comment":"The inference-time assignment rule for MaskCentroid is never defined. Equations (2) and (3) describe a centroid F' and a similarity phi_j(e_i) using the known set N_j of pixels belonging to an instance, but the paper does not specify how a predicted high-confidence keypoint anchors pixels, how the similarity threshold or assignment procedure works, how overlapping instances are separated, or what happens when the anchoring keypoint is occluded or incorrectly detected. The text 'defining the highly confident keypoint as a learnable parameter' is not an algorithm. The training loss is also not written: the paper only says 'the MaskCentroid offset loss' is minimized (Section III). Without this specification, the segmentation pipeline cannot be reproduced or even falsified from the manuscript, which is a load-bearing gap for the central claim that dynamic MaskCentroids improve mask AP.","section":"Section III-D (Dynamic MaskCentroid)"},{"comment":"The key ablation comparing Static MaskCentroid (SMc) and Dynamic MaskCentroid (DMc) is presented only as a figure with no numerical values. The text states that 'the exceptional performance of the proposed DMc approach demonstrates its effectiveness' but provides no AP numbers, no standard deviation, and no evaluation protocol details for Fig. 6. Since the dynamic-centroid design is the paper's main novelty, this ablation must be reported with concrete numbers (e.g., a table with AP, AP_50, AP_75) for the claim to be verifiable.","section":"Section IV-A, Fig. 6"},{"comment":"The abstract claims 'real time performance advantages' and 'execution frame rate per second' advantages, and Section IV states that Fig. 5 shows 'high FPS' relative to sister models. However, no FPS table or per-model frame-rate numbers are given anywhere in the manuscript. Fig. 5 is a scatter plot without a clear quantitative legend. A real-time claim requires at least a table reporting FPS, hardware, batch size, and input resolution for each compared model; the current evidence is insufficient to assess the central speed claim.","section":"Abstract and Section IV (Computational Cost)"}],"minor_comments":[{"comment":"The text says VISUALCENT 'improves 10.0% compare to HGG [18]' using the OCHuman test set; since the improvement is from 36.0 to 46.0 AP, this is a 10.0 percentage-point gain, not a 10.0% relative improvement. Please clarify the terminology.","section":"Section IV (OCHuman results)"},{"comment":"Table III lists PersonLab four times, with two rows per backbone and identical backbone labels but different AP values. Please label which rows are single-scale and which are multi-scale, and reconcile the caption footnote with the table entries.","section":"Table III"},{"comment":"In the Segmentation Results paragraph, 'Pose+Seg [9]' appears to be a citation error: reference [9] is Dantone et al., not PosePlusSeg; the PosePlusSeg reference is [1]. Similarly, in Section IV the MIPNet citation appears as [19] while the reference list shows MIPNet as [20].","section":"References"},{"comment":"The method name is spelled inconsistently: 'VISUALCENT' in the abstract and 'VISUAL CENT' in the body and title. Also, 'MaskCenroid' appears in Section I. Please unify the spelling.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the missing MaskCentroid specification is well-founded and is the main obstacle. The paper's reported COCO and OCHuman numbers are internally consistent in the tables, but the central mechanism is not described with enough precision to be checked. This is fixable in a revision: the authors should provide the full inference-time clustering algorithm, the loss formula for the MaskCentroid offset, and a quantitative SMc-vs-DMc ablation. The real-time claim also needs an FPS table. I see no evidence of circular reasoning; the evaluation is on external benchmarks. If the requested material is provided and the claims remain supported, the paper could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the pose numbers are strong, the segmentation mechanism is not specified well enough to check, and the paper's headline claims would need a serious rewrite before they're believable.\n\nWhat's actually new: KeyCentroid is a disk-constrained keypoint offset regression, a plausible variation on PersonLab's short-range offsets and CenterNet's center offsets. MaskCentroid is the more interesting idea: anchor mask clustering at a high-confidence keypoint so the anchor can move dynamically. The COCO keypoint results (76.1 AP with ResNet-152) and the OCHuman numbers are the kind of numbers that would move this subfield if they reproduce.\n\nWhat the paper does well: it evaluates on public benchmarks, compares directly against a reasonable list of bottom-up baselines including its own PosePlusSeg, and the KHDR+-KeyCentroid ablation gives concrete numbers. That part is checkable and looks sincere.\n\nSoft spots, in order. Biggest: the MaskCentroid clustering rule is never actually specified. Equations (2) and (3) define a mean pixel position and a Gaussian similarity to a mean embedding, but both formulas assume the instance N_j is already known. There is no inference-time assignment rule that takes a predicted high-confidence keypoint and produces a set of pixels for that instance, and no training loss is written for the mask centroid offset. Saying the keypoint is a learnable parameter is not an algorithm. So the central segmentation claim cannot be reproduced or even falsified from the manuscript. The stress-test note is right, and it's not a minor omission.\n\nNext: numeric inconsistencies. The \"10.0% improvement\" over HGG is an absolute 10.0 AP point gap, which at the reported values is around 28% relative. The \"3.5%\" over MIPNet is likewise a point gap. Table III labels Pose+Seg as [1] in the table but [9] in the text, and reference [9] is Dantone et al., not PosePlusSeg. The reference list also duplicates [5]. These are likely carelessness, but they erode confidence in the reported deltas.\n\nThird: the dynamic-vs-static MaskCentroid comparison is a figure with no numbers, and the FPS/real-time claim is a plot with no table. For a paper that leads with a real-time claim, that's under-evidenced.\n\nBottom line: the pose side is salvageable and worth a serious referee. The segmentation side, as written, is not verifiable. I would send it to review only with the expectation that the authors supply the assignment rule, the mask loss, and numeric ablations. It deserves referee time rather than desk rejection because the reported results, if real, are a clear within-subfield improvement and the system idea is coherent. For me personally, I'd cite the KeyCentroid pose result with caution; I wouldn't cite the MaskCentroid mechanism until it's written down.\n\nWho it's for: bottom-up pose/segmentation researchers, especially those working on crowded real-time scenarios. They'd get something from the disk-offset idea; they'd need to wait for a revised version on the clustering.\n\nRecommendation: conditional acceptance after major revision, with the missing algorithm and data supplied.","headline":"Reported bottom-up pose and segmentation numbers are strong, but the MaskCentroid grouping rule is never specified, so the central segmentation claim cannot be checked from the paper.","tokens_in":9574,"tokens_out":3867,"would_cite":false,"duration_ms":36696,"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":"VISUALCENT combines keypoint detection and instance segmentation in one bottom-up network using dynamic keypoint centroids as grouping anchors.","keywords":["human pose estimation","instance segmentation","bottom-up","dynamic centroid","keypoint heatmap","occlusion","COCO","OCHuman"],"falsifier":"Re-run the static-versus-dynamic MaskCentroid comparison on the COCO validation set and on OCHuman with numeric mask AP, replacing the dynamic keypoint anchor with a fixed geometric centroid (the mean of the person's predicted keypoints or the static centroid used in PosePlusSeg). If the static variant matches or exceeds the dynamic variant's AP, then the dynamic anchor is not the source of the reported improvement.","tokens_in":8620,"feed_emoji":"🧍","tokens_out":9512,"duration_ms":84030,"temperature":0.7,"pith_summary":"VISUALCENT is a unified bottom-up framework for multi-person human pose estimation and instance-level segmentation that does not need a person detector. The paper's central idea is to use a high-confidence predicted keypoint as a dynamic centroid (MaskCentroid) to which all mask pixels of that person are attached by embedding distance, so grouping is cheap and adjusts to movement and occlusion. The authors report that this design outperforms prior unified models such as PosePlusSeg and PersonLab on COCO and OCHuman benchmarks, and runs at a higher frame rate. The practical payoff is a scalable, real-time visual human analysis system for crowded or entangled scenes.","feed_headline":"Dynamic keypoint anchors unify pose and mask in one net","feed_subtitle":"VISUALCENT uses a high-confidence joint as a moving anchor, beating prior bottom-up models on COCO and OCHuman.","key_machinery":"The load-bearing mechanism is the dynamic centroid, used in two linked operations. KeyCentroid is a two-channel regression within the keypoint disk (radius $R=32$) that points toward the joint's location, refining the coarse heatmap response and providing high-confidence keypoint coordinates. MaskCentroid then selects one of those high-confidence keypoints per person as a learnable center of attraction for pixel clustering, grouping pixels by embedding distance to that center with the exponential affinity in equation (3). The dynamic adjustment of the centroid is what the paper credits for keeping segmentation stable under rapid body motion and severe occlusion, and it is also what keeps the runtime low, since each pixel only compares against a few anchors rather than all other pixels.","core_discovery":"The core claim is that a single dynamic centroid per person, derived from a confidently detected keypoint, resolves the pixel-to-instance grouping problem in bottom-up segmentation. VISUALCENT first produces keypoint heatmaps with a disk representation (a radius-32 region around each joint) and a KeyCentroid regression that points from any pixel in the disk to the joint's exact location; these two outputs are combined to pick an optimal keypoint. One such keypoint per instance is then treated as a learnable MaskCentroid, and all mask pixels are clustered by embedding proximity using $\\phi_j(e_i)=\\exp(-\\lVert e_i-\\mu_j\\rVert^2 / 2\\sigma_j^2)$, where $\\mu_j$ is the mean embedding of the instance. The paper claims this dynamic choice beats a fixed centroid when bodies move or occlude each other, and on COCO test-dev reaches 76.1 keypoint AP with ResNet-152 and 47.6 mask AP on the segmentation test set, along with gains on OCHuman.","pith_inferences":["The MaskCentroid idea is portable to other instance-level dense tasks, such as animal pose and part segmentation, whenever one reliable semantic keypoint per instance exists.","If the anchor is the single highest-confidence keypoint, segmentation accuracy becomes coupled to keypoint confidence calibration; a separate learned anchor-confidence could decouple them.","A quantitative ablation that varies the choice of anchor keypoint (e.g., highest confidence vs. a limb-specific joint) would clarify how much of the gain comes from the dynamic selection rule itself.","The method's limitation is likely a person with no visible keypoint at all; adding a fallback centroid, such as the mean of lower-confidence keypoints or a learned embedding center, would extend it to fully occluded instances."],"forward_implications":["A single bottom-up model could replace the detect-then-segment cascade for human analysis, eliminating the extra compute and failure modes of a person detector.","The per-pixel grouping cost scales with the number of anchors, so crowded frames remain tractable without pairwise pixel comparisons.","Real-time video applications such as surveillance, human-robot interaction, and live sports analysis could run pose and mask estimation at interactive frame rates.","The dynamic-centroid mechanism appears to transfer across backbone capacities, since the paper shows gains with both ResNet-101 and ResNet-152."],"supporting_citations":[{"why":"It supplies the direct predecessor baseline and the static-centroid grouping approach that MaskCentroid is designed to improve; its COCO numbers are the main comparative points.","marker":"[1]"},{"why":"It is the closest prior bottom-up joint pose and instance segmentation model, providing the geometric-embedding baseline and multi-scale results VISUALCENT compares against.","marker":"[11]"},{"why":"It is the standard top-down detect-then-segment baseline whose segmentation accuracy and computational cost the paper contrasts with its detector-free approach.","marker":"[14]"},{"why":"It provides the detection-free segmentation baseline on OCHuman and introduced the OCHuman dataset used for the occlusion-heavy evaluation.","marker":"[46]"},{"why":"It supplies the HGG baseline numbers on the OCHuman keypoint validation and test sets that VISUALCENT reports improving upon.","marker":"[18]"},{"why":"It supplies the MIPNet baseline numbers on the OCHuman keypoint validation and test sets that VISUALCENT reports improving upon.","marker":"[20]"}],"fun_headline_variants":["Dynamic centroid unifies pose and mask in real time","One net for pose and mask via moving keypoint anchor","Keypoint-driven centroid boosts bottom-up pose and segmentation","VISUALCENT: dynamic centroid for fast pose and instance masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that at least one keypoint per person is detected with high confidence even when bodies overlap or are occluded, because that single keypoint is the anchor that gathers all of the person's mask pixels; if no keypoint is reliable or the chosen one belongs to a different instance, the segmentation grouping has no correct center to cluster around.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic centroid unifies pose and mask in real time","One net for pose and mask via moving keypoint anchor","Keypoint-driven centroid boosts bottom-up pose and segmentation","VISUALCENT: dynamic centroid for fast pose and instance masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000155,"raw_usage":{"total_tokens":1181,"prompt_tokens":877,"completion_tokens":304,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":493,"tokens_out":304,"duration_ms":3601,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:03:27.128934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the static-versus-dynamic MaskCentroid comparison on the COCO validation set and on OCHuman with numeric mask AP, replacing the dynamic keypoint anchor with a fixed geometric centroid (the mean of the person's predicted keypoints or the static centroid used in PosePlusSeg). If the static variant matches or exceeds the dynamic variant's AP, then the dynamic anchor is not the source of the reported improvement.","supporting_citations":[{"cited_title":"Ahmad, J","cited_arxiv_id":null,"evidence_quote":"It supplies the direct predecessor baseline and the static-centroid grouping approach that MaskCentroid is designed to improve; its COCO numbers are the main comparative points."},{"cited_title":"Gidaris, Tompson, and K","cited_arxiv_id":null,"evidence_quote":"It is the closest prior bottom-up joint pose and instance segmentation model, providing the geometric-embedding baseline and multi-scale results VISUALCENT compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is the standard top-down detect-then-segment baseline whose segmentation accuracy and computational cost the paper contrasts with its detector-free approach."},{"cited_title":"Zhang, R","cited_arxiv_id":null,"evidence_quote":"It provides the detection-free segmentation baseline on OCHuman and introduced the OCHuman dataset used for the occlusion-heavy evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the HGG baseline numbers on the OCHuman keypoint validation and test sets that VISUALCENT reports improving upon."},{"cited_title":"Khirodkar, V","cited_arxiv_id":null,"evidence_quote":"It supplies the MIPNet baseline numbers on the OCHuman keypoint validation and test sets that VISUALCENT reports improving upon."}],"review_version":1}