{"id":"157ab667-ae33-446f-adae-196463c7cc52","arxiv_id":"2411.15653","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"OCDet predicts object center heatmaps with Generalized Centerness and Balanced Continuous Focal Loss, and reports higher recall and CAS than YOLO11 on edge NPUs.","lead":"OCDet detects object centers by predicting bounding box-aware heatmaps and extracting peaks. It reports higher center detection recall than YOLO11 on COCO with fewer parameters, less computation, and lower NPU latency, on a metric the authors introduce.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (9) defines CAS with an undefined denominator N_i and an unaveraged MD_i, so the headline CAS comparisons are not reproducible as written.","rationale":"The paper is a substantial engineering contribution with a coherent pipeline: GC generalizes FCOS centerness to heatmap generation, BCFL is a reasonable adaptation of QFL with target-dependent weighting, and the ablations are internally consistent. I read the empirical claims in good faith, and the code release is a point in the paper's favor. The single most load-bearing condition for the central claim is that CAS is a well-defined, normalized metric, because every quantitative comparison in the abstract, Figure 1, and Table 1 is expressed in CAS units. The reader's weakest assumption identifies exactly this condition, and I agree with that diagnosis. Eq. (7) calls MD_i an average but writes only a sum, and Eq. (9) divides by N_i without defining it. These are internal specification gaps, not disagreements with an external consensus. The values in Table 1 are plausible under more than one normalization, so they do not disambiguate the formula. This is unlike the missing CenterNet comparison or the validation-set tuning, which are genuine but secondary concerns: they affect the strength of the generalization claim, whereas an ill-defined metric blocks verification of the headline numbers themselves. A code-level check can settle the issue. If the code confirms a sensible normalization and reproduces Table 1, the paper's claims should be accepted in the conditional sense the reader already adopted; if not, the quantitative claims must be revised. Therefore the reader's CONDITIONAL verdict remains the right one, and my stress-test does not change it.","tokens_in":21172,"tokens_out":5600,"duration_ms":57602,"concrete_test":"Clone the released repository and run the evaluation script on COCO val2017 for the OCDet-N and YOLO11n settings. Inspect the exact code path for Eqs. (8)–(9) and record (a) how CP_i is normalized (by number of images, by number of GT objects, or by another count) and (b) whether MD_i is divided by the number of matched pairs before averaging. Recompute the two CAS values from raw matching outputs. If the code's normalization differs from a literal reading of Eq. (9), or if either CAS value shifts by more than 0.01, the headline claim is not reproducible as written. If no evaluation code is available, the authors should be asked to provide it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—up to 23% higher CAS with fewer parameters, less computation, and lower latency—rests entirely on the proposed CAS metric, and Eq. (9) is not well defined as printed. Eq. (7) states that MD_i is the average of the normalized distances but writes only a sum, omitting the division by |M|. Eq. (9) then subtracts ΣCP_i/N_i and ΣMD_i/N_i without ever defining N_i. If N_i is the number of images, the CP term is the mean raw unmatched-point count per image, which is not bounded and does not fit the 'normalized distance metric' framing; if N_i is instead a per-image object count, that normalization is never stated, and CP_i = max(#unmatched preds, #unmatched GT) can still exceed N_i. No derivation, pseudocode, or code excerpt for the metric is included, so the GitHub link alone does not pin down the formula used to produce Table 1. Because every headline number is expressed in CAS units, this ambiguity is load-bearing, not cosmetic. The concern is not that the method is wrong; the reported improvements over YOLO11 and the KP baselines simply cannot be checked until the metric is fully specified and ideally reproduced from the released code.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces OCDet, a heatmap-based object center detection framework for NPU-equipped edge devices. Ground-truth heatmaps are generated from bounding-box annotations via a Generalized Centerness (GC) map with shape parameters eta and phi; the models use MobileNetV4 backbones with Semantic FPN and are trained with a Balanced Continuous Focal Loss (BCFL). Evaluation is performed with a proposed Center Alignment Score (CAS) that uses Hungarian matching, a matched-discrepancy term, and a cardinality penalty. On COCO, the authors report that OCDet variants achieve up to 23% higher CAS than YOLO11 while using 42% fewer parameters, 34% less computation, and 64% lower NPU latency, and up to 186% higher CAS than identical models trained under a standard keypoint-detection framework. The paper includes ablations over ground-truth type, GC exponents, loss variants, BCFL hyperparameters, backbone, architecture, output stride, FPN modules, and peak-identification settings.","tokens_in":21367,"tokens_out":5841,"duration_ms":49303,"significance":"If the quantitative claims survive scrutiny, this is a practically useful contribution: it defines a cheap edge-deployable center-detection task, proposes a box-aware heatmap representation, and provides extensive ablations, NPU latency measurements, and released code. The strengths are real: the architecture and loss design are reasonable, the ablations are informative, and the qualitative failure cases are honestly discussed. However, the central quantitative claims are currently not verifiable because the CAS metric is incompletely specified, and the evaluation protocol has selection-on-validation issues. The significance of the specific numerical advantages over YOLO11 and the keypoint baselines is contingent on fixing the metric definition and adding held-out evaluation with variance estimates.","major_comments":[{"comment":"Eq. (7) defines MD_i as the 'average' of normalized matched distances but writes only a sum, omitting the division by |M|. Eq. (9) then subtracts sum CP_i/N_i and sum MD_i/N_i without ever defining N_i. If N_i is the number of images, the CP term is a mean raw unmatched-point count, which is neither normalized nor bounded and would make the reported CAS values in Tab. 1 (0.2-0.4) impossible for typical COCO image object counts; if N_i is instead an object count or a per-image normalization, that definition is absent. The text also provides no pseudocode or code excerpt for the metric, so the GitHub link alone does not pin down the formula used to produce Table 1. Because every headline comparison—including the 23% CAS improvement over YOLO11 and the 186% over the KP baselines—is expressed in CAS units, the central claim cannot be checked as written. Please specify N_i, the averaging in MD_i (including the case |M|=0), and the exact per-image normalization, and verify that the released code reproduces Table 1.","section":"Sec. 3.5, Eqs. (7)-(9)"},{"comment":"The validation set val2017 is used both to select hyperparameters and to report final results. Specifically, Tab. 5 selects the probability threshold (0.6) and alpha for BCFL on the person center-detection task using CAS, Tab. 6 selects gamma, and Appendix E selects the peak-identification probability threshold (0.5) and minimum distance (0.3) on OCDet-M using CAS; these choices are then applied across all models in Tab. 1. This creates a selection loop: the same data and metric that define the reported superiority are used to tune the method. At minimum, report results on a held-out split or use nested cross-validation, and quantify the selection bias. Without this, the 'up to 23%' claim may be an optimistic estimate.","section":"Sec. 4.1, Sec. 4.3, and Appendix E"},{"comment":"The comparison with YOLOv8 and YOLO11 extracts centers from bounding-box outputs of off-the-shelf detectors, but no confidence threshold, NMS setting, or other center-extraction protocol is reported for these baselines, and the detectors are not trained or adapted for the center-detection objective. YOLO models are optimized for box IoU, not center alignment, so the comparison conflates task definition with model capability. Since the claimed Pareto dominance in Fig. 1 depends on this protocol, please provide a sensitivity analysis over YOLO confidence thresholds and report the exact thresholds used, or explicitly reframe the claims as 'center extraction from off-the-shelf detectors' rather than 'OCDet outperforms YOLO11 in center detection.'","section":"Table 1 and Sec. 4.2"},{"comment":"No error bars or multiple-run statistics are reported. Several headline differences are small in absolute terms (e.g., OCDet-S 0.313 vs YOLO11s 0.309; OCDet-M 0.362 vs YOLOv8m 0.357), and without seed variance it is impossible to determine whether the 1% and 6% improvements are significant. Please report mean +/- std over at least three seeds for the main comparison table and for the key ablations (Tabs. 1, 2, 4, 5, 6), and state the number of runs used.","section":"Tables 1-10"}],"minor_comments":[{"comment":"MD_i is called an average but the division by |M| is missing; please add the denominator and handle the case |M|=0 explicitly.","section":"Sec. 3.5, Eq. (7)"},{"comment":"The summation indices and the definition of N_i are ambiguous; use consistent notation, for example CAS = 1 - (1/I) * sum_i (CP_i + MD_i), and define all quantities.","section":"Sec. 3.5, Eq. (9)"},{"comment":"The phrase 'Matched Distance (MD) and Coverage Precision (CP)' should be 'Matched Discrepancy and Cardinality Penalty' to match Eq. (8) and the rest of the paper.","section":"Sec. 3.5, text"},{"comment":"The permutation notation sigma in Sigma_N with the assumption N > G is only valid when predictions outnumber ground truths; if G > N, the matching should be defined over injections. Please state the general case.","section":"Sec. 3.4, Eq. (5)"},{"comment":"The ablation tables (Tabs. 2-6 and 7-10) are for person center detection only, but the selected hyperparameters are then applied to the 80-class COCO task; please justify the transferability or provide 80-class ablations.","section":"Sec. 4.3 and Appendix E"},{"comment":"The '2x schedule' with 24 epochs is nonstandard; please clarify the schedule definition.","section":"Sec. 4.1 and Appendix D"},{"comment":"Reference [56] for FCOS appears to point to a paper on 3D object detection on LiDAR range images; the correct FCOS reference (Tian et al., ICCV 2019) should be cited.","section":"References"},{"comment":"There are typos: 'CAS grains' should be 'CAS gains', and the final sentence of Sec. 4.2 contains a comma splice; please proofread.","section":"Sec. 1 and Sec. 4.2"},{"comment":"The 'probability threshold' of 0.6 used to derive alpha and the peak-identification probability threshold of 0.5 are different concepts but are both called 'probability threshold'; please use distinct terms to avoid confusion.","section":"Sec. 4.1 and Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely a better fit for an applied edge-AI venue than for a purely methodological venue; the novelty of the metric and loss is modest and the main empirical claims are not yet reproducible. The central ideas are plausible, but the undefined CAS normalization and the selection-on-validation protocol are load-bearing and must be fixed before the quantitative claims can be assessed. I would not reject on novelty alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The engineering here is real, and the paper earns a serious look. The authors build a lightweight heatmap-based center detector for NPUs, measure actual latency on i.MX 8M Plus across five model sizes, and give thorough ablations on ground-truth heatmap type, loss, backbone, architecture, and peak-threshold settings. The code is released. That is a solid applied contribution, especially for someone working on edge surveillance or industrial vision.\n\nWhat is new is modest but genuine. Generalized Centerness is a direct generalization of FCOS's Centerness, and the default eta=phi=0.5 makes the default heatmap exactly FCOS Centerness; the novelty is in using it as a dense ground-truth heatmap rather than a weighting factor. BCFL is QFL plus a target-dependent alpha weight. That is an incremental but useful combination, and the ablations show it helps, particularly for small objects. The CAS metric is new, and the Hungarian-matching framing is reasonable.\n\nThe soft spots are real, and one is load-bearing. As printed, Eq. (7) calls MD_i an \"average\" but writes only a sum with no division by |M|, and Eq. (9) subtracts sums over an undefined N_i. The stress-test note claims the resulting CAS values are \"impossible\" if N_i is the number of images; I think that part is overstated, since an average of ~0.6 unmatched points per image could produce the reported CP values. But the deeper point stands: the normalization is not defined, the paper gives no pseudocode or code excerpt for the metric, and the released GitHub link does not pin down the formula in the text. Every headline comparison in the abstract and Table 1 is expressed in CAS units, so this ambiguity blocks independent verification of the central claim. That is a major revision issue, not a desk-reject issue.\n\nOther problems are more conventional but worth naming: hyperparameters are tuned on the validation set and the same set is used for reporting; no error bars are given; and the most natural baseline, CenterNet, is cited but never compared. The YOLO baselines are trained for box detection, so comparing them on center extraction is a bit apples-to-oranges, though the authors do acknowledge this in the discussion. None of these are fatal, but together they mean the empirical claims are not yet fully pinned down.\n\nWho should read this: practitioners doing coarse object localization on NPU-equipped edge devices. It is not a field-reshaping result, but it is a useful, honest engineering paper with good latency data. I would send it to peer review, but the reviewers should require a fully specified CAS formula (with division and N_i made explicit) and ideally a reproduction from the released code, plus a CenterNet comparison and either a held-out split or error bars. The core method is probably sound; the metric definition has to be fixed before the headline numbers can be trusted.","headline":"Genuinely practical edge-object-centering work with real NPU latency data, but the headline CAS numbers rest on a metric that is not fully specified as printed.","tokens_in":21983,"tokens_out":2393,"would_cite":false,"duration_ms":24768,"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":"Bounding-box-aware center heatmaps beat YOLO11 for coarse object localization on edge NPUs, this paper argues.","keywords":["object center detection","heatmap prediction","edge NPU","Generalized Centerness","Balanced Continuous Focal Loss","Center Alignment Score","MobileNetV4","Semantic FPN"],"falsifier":"Run the released code on the COCO val2017 set, compute Eq. (9) with N_i interpreted as the number of images, and check whether the resulting per-image penalties (CP around 0.5-0.65 for OCDet variants in Tab. 1) are consistent with typical object counts per image (roughly 7-10); if the metric instead normalizes by object count or matched pairs, the tabulated CAS values change and the headline percentages against YOLO11 need recomputation. A direct way to settle it: reproduce a single row of Tab. 1 from the public code and compare the computed CAS to the printed 0.297.","tokens_in":20896,"feed_emoji":"🎯","tokens_out":6412,"duration_ms":52520,"temperature":0.7,"pith_summary":"OCDet claims that for coarse object localization on edge devices with NPUs, predicting object center heatmaps is more effective than full bounding-box detection or keypoint detection. The framework combines Generalized Centerness (GC), a bounding-box-derived probability map, with Balanced Continuous Focal Loss (BCFL) and a lightweight Semantic FPN with MobileNetV4 backbones. Under the paper's Center Alignment Score (CAS) metric, OCDet is reported to exceed YOLO11 by up to 23% while using 42% fewer parameters, 34% less computation, and 64% lower NPU latency, and to beat a keypoint-detection baseline by up to 186% with identical models. If these numbers hold, surveillance, industrial automation, and other high-recall edge applications could get more reliable coarse localization at lower hardware cost.","feed_headline":"Center heatmaps beat YOLO11 on edge NPUs","feed_subtitle":"Bounding-box-aware heatmaps plus a balanced focal loss: 23% higher CAS, 64% lower latency.","key_machinery":"The load-bearing object is Generalized Centerness (GC), the identity GC(x,y) = (min(l,r)/max(l,r))^eta times (min(t,b)/max(t,b))^phi, which converts a bounding box into a [0,1] per-pixel probability heatmap whose concentration is tuned by eta and $\\varphi$ (eta=phi=0.5 reproduces the Centerness of FCOS). It replaces the fixed-Gaussian ground truth of keypoint frameworks. Two supporting mechanisms carry the training and evaluation: Balanced Continuous Focal Loss (BCFL), which reweights Quality Focal Loss by an adaptive factor alpha_c(y) = $\\alpha$*y + (1-$\\alpha$)*(1-y) to suppress low-value background points and emphasize high-value center regions; and the Center Alignment Score (CAS), which applies Hungarian matching and subtracts a Matched Discrepancy and a Cardinality Penalty from 1. The models are built on Semantic FPN with MobileNetV4, chosen for NPU-friendly operations.","core_discovery":"On the paper's own terms, the discovery is that the standard practice of placing a fixed Gaussian at each target point, and the standard practice of regressing bounding boxes, are both mismatched to the object-center detection task. OCDet replaces them with Generalized Centerness, which assigns every pixel inside a box a probability of being the center based on the ratio of distances to the box edges, with shape controlled by hyperparameters eta and phi. Under this supervision, models trained with the proposed Balanced Continuous Focal Loss learn heatmaps that better preserve central geometry, and the paper reports that this combination achieves consistently higher Center Alignment Scores than YOLO11 and YOLOv8 across all five model variants, with substantially smaller models and lower latency, and higher recall and F1.","pith_inferences":["A testable extension would be to attach GC ground truth to an existing CenterNet-style detector and compare against the fixed-Gaussian variant, isolating whether the reported gains come from the heatmap shape or from the FPN+BCFL combination.","A scale-free version of CAS could divide the cardinality penalty by the per-image ground-truth count; such normalization would remove the N_i ambiguity and make cross-dataset comparisons safer.","Learning eta and phi per class is a natural next step: the ablation's finding that skis prefer eta=0.9/phi=0.1 and bottles prefer eta=0.1/phi=0.9 suggests a per-class shape model could recover the small-object CASS deficit.","On the hardware side, the paper's exclusion of SE blocks and LayerNorm because of NPU latency is consistent with a simple latency-overhead intuition; a direct per-operation timing measurement on the i.MX 8M Plus would confirm how much of the 64% latency win is architectural versus quantization-related."],"forward_implications":["If CAS is the right yardstick, then object center detection on COCO-like data is better served by heatmap regression with GC ground truth than by YOLO-family box regression, at least under the reported protocol: up to 23% higher CAS with 42% fewer parameters and 64% lower NPU latency.","Replacing the fixed Gaussian ground truth of keypoint frameworks with GC is predicted to raise CAS by up to 186% at identical model size, with the largest relative gains at the smallest model scale (OCDet-N versus KP-N).","The ablation results imply that BCFL is the component that unlocks small-object center detection: CASS jumps from 0.044 (QFL) to 0.314 (BCFL) in the loss comparison table.","Because GC only needs bounding boxes, the framework transfers to any object detection dataset and to automated box-labeling pipelines without extra manual keypoint annotation.","A continuous, scale-aware metric like CAS could complement or replace threshold-based AP for center-focused localization tasks, since it avoids the abrupt performance changes induced by fixed IoU thresholds."],"supporting_citations":[{"why":"Supplies the Centerness concept that GC generalizes into a full heatmap generator.","marker":"[56]"},{"why":"Provides Quality Focal Loss, which BCFL adapts by adding the alpha_c(y) target-dependent weighting.","marker":"[33]"},{"why":"Contributes the focal-loss modulating factor and hard-negative mining that BCFL inherits.","marker":"[49]"},{"why":"Defines the MobileNetV4 backbone family whose NPU-friendly blocks all OCDet variants build on.","marker":"[43]"},{"why":"Supplies the Semantic FPN upsampling-and-summing head architecture used by the OCDet models.","marker":"[28]"},{"why":"Provides the Hungarian matching algorithm used to pair predicted and ground-truth centers for CAS.","marker":"[30]"},{"why":"Is the COCO dataset and AP-style evaluation protocol on which all experiments and comparisons are based.","marker":"[35]"},{"why":"Is the primary real-time detector baseline whose center outputs OCDet is claimed to outperform.","marker":"[27]"},{"why":"Is the second YOLO baseline used in the five-variant Pareto comparison.","marker":"[26]"}],"fun_headline_variants":["Heatmap method beats YOLO11 on edge NPUs with 64% less latency","OCDet: 23% better center detection, 64% faster on NPUs","Bounding-box heatmaps outclass YOLO11 on edge AI chips","Smaller model, faster speed: OCDet wins center detection on NPUs","Generalized centerness heatmaps: up to 186% gain over keypoint nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire comparison rests on the Center Alignment Score being a well-defined, identically normalized metric across all models; Eq. (9) of the paper sums CP_i/N_i and MD_i/N_i without ever defining the count N_i, so the reported CAS values are only meaningful if that normalization is fixed in a way the paper does not state.","fun_headline_variants_meta":{"raw":{"variants":["Heatmap method beats YOLO11 on edge NPUs with 64% less latency","OCDet: 23% better center detection, 64% faster on NPUs","Bounding-box heatmaps outclass YOLO11 on edge AI chips","Smaller model, faster speed: OCDet wins center detection on NPUs","Generalized centerness heatmaps: up to 186% gain over keypoint nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000372,"raw_usage":{"total_tokens":2008,"prompt_tokens":986,"completion_tokens":1022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":914}},"tokens_in":602,"tokens_out":1022,"duration_ms":7186,"temperature":1.0,"reasoning_tokens":914,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:03:47.216432+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on the COCO val2017 set, compute Eq. (9) with N_i interpreted as the number of images, and check whether the resulting per-image penalties (CP around 0.5-0.65 for OCDet variants in Tab. 1) are consistent with typical object counts per image (roughly 7-10); if the metric instead normalizes by object count or matched pairs, the tabulated CAS values change and the headline percentages against YOLO11 need recomputation. A direct way to settle it: reproduce a single row of Tab. 1 from the public code and compare the computed CAS to the printed 0.297.","supporting_citations":[{"cited_title":"Fully convolutional one-stage 3d object detection on lidar range images","cited_arxiv_id":null,"evidence_quote":"Supplies the Centerness concept that GC generalizes into a full heatmap generator."},{"cited_title":"Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection","cited_arxiv_id":null,"evidence_quote":"Provides Quality Focal Loss, which BCFL adapts by adding the alpha_c(y) target-dependent weighting."},{"cited_title":"Focal loss for dense object detection","cited_arxiv_id":null,"evidence_quote":"Contributes the focal-loss modulating factor and hard-negative mining that BCFL inherits."},{"cited_title":"Panoptic feature pyramid networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Semantic FPN upsampling-and-summing head architecture used by the OCDet models."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"Is the COCO dataset and AP-style evaluation protocol on which all experiments and comparisons are based."},{"cited_title":"Ultralytics yolo11","cited_arxiv_id":null,"evidence_quote":"Is the primary real-time detector baseline whose center outputs OCDet is claimed to outperform."},{"cited_title":"Ultralytics yolov8","cited_arxiv_id":null,"evidence_quote":"Is the second YOLO baseline used in the five-variant Pareto comparison."}],"review_version":1}