{"id":"24b113cb-d7d1-44a0-a706-28fd0bda08aa","arxiv_id":"2412.01422","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MamKPD, a Mamba-based 2D keypoint detector with a contextual modeling module, reports 77.3% AP on COCO at 1492 FPS and top MPII accuracy.","lead":"MamKPD is a new Mamba-based neural network for real-time 2D keypoint detection, reporting 77.3% AP on COCO at 1492 FPS. It adds a small convolutional module to the Mamba backbone to improve context modeling and is presented as the first Mamba design for this task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Speed claim rests on unmatched FPS measurement: top-down pose-branch-only numbers are compared to one-stage full-pipeline baselines and A100-reported ViTPose speeds, so the headline speed advantage is not yet established.","rationale":"The reader's verdict is CONDITIONAL, and my read supports keeping that verdict rather than moving to accept or reject. The reader's weakest assumption was comparison fairness, focusing on the A100-versus-4090 speed mismatch and the no-pretraining MPII baselines. I agree with that diagnosis and sharpen it: the most damaging issue is that MamKPD is top-down and its FPS is being compared against one-stage full-pipeline methods. A top-down detector is mandatory before the keypoint branch runs, so the reported 1492 FPS cannot be the practical throughput of the system. This is the single most load-bearing concern because the abstract's central quantitative claim is the speed-accuracy tradeoff: '77.3% AP ... with 1492 FPS.' If the speed comparison is not apples-to-apples, the paper's main practical conclusion is unsupported. The architecture itself may still be a useful Mamba-based pose backbone; the ablations show the CMM and stem contribute meaningfully, and the no-pretraining COCO results are interesting. The paper is a plausible contribution, but the empirical headline needs a matched benchmark before it can be considered established. My recommendation therefore remains CONDITIONAL, equivalent to UNCHANGED, pending code release and a fair speed measurement protocol.","tokens_in":15130,"tokens_out":5366,"duration_ms":49303,"concrete_test":"Benchmark MamKPD-L, RTMO-L, KAPAO-L, and ViTPose-B on the same RTX 4090 with identical settings: same input resolution (256x192 or the resolution used by each baseline), batch size 1, same FP16/FP32 precision, and the same CUDA version. Measure (i) MamKPD-L pose-branch-only FPS on COCO val crops, (ii) MamKPD-L end-to-end FPS including its person detector, and (iii) RTMO-L and KAPAO-L end-to-end FPS on the same COCO val images. Report AP and FPS for all settings. If MamKPD-L end-to-end FPS is close to or below RTMO-L, the claimed speed superiority collapses; if it stays near 1492 FPS end-to-end, the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim is that MamKPD is dramatically faster than prior pose estimators while retaining accuracy. That claim depends entirely on the FPS column of Table 2, and that column is not a matched comparison. MamKPD is a top-down method: the text states that a person detector must run first, yet the reported 1492 FPS appears to measure only the keypoint branch on cropped instances. RTMO and KAPAO are one-stage methods whose quoted FPS includes the detection and pose estimation in a single forward pass. Comparing a pose-branch-only number to their full-pipeline numbers overstates the real-time advantage. The paper itself admits in Section 4.2.2 that ViTPose speeds were measured on an A100 while MamKPD was measured on an RTX 4090. Input resolution, batch size, FP16/FP32 precision, and whether postprocessing or detector time is excluded are not reported. If end-to-end FPS were measured, MamKPD's margin over RTMO and KAPAO could shrink substantially or disappear. The MPII comparison has a related fairness problem: ViTPose and HRNet are reproduced without ImageNet pretraining in Table 3, while MamKPD is also trained from scratch, so the 'state-of-the-art on MPII' claim compares against weakened baselines. Neither issue invalidates the architecture, but both must be resolved before the headline speed and SOTA claims can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents MamKPD, a top-down 2D keypoint detector built on a Mamba/SSM backbone. The architecture adds a Context Modeling Module (CMM) before the SS2D blocks in each of three stages, using depthwise convolutions for inter-patch dependencies and linear/1x1 depthwise convolutions for intra-patch feature extraction, followed by a simple deconvolutional decoder. Three model sizes are defined and evaluated on COCO val2017, MPII validation, and AP-10K. The main claims are that MamKPD is the first Mamba-based 2D keypoint detection architecture, that MamKPD-L reaches 77.3 AP on COCO with 1492 FPS on an RTX 4090, that it is state-of-the-art on MPII, and that it is competitive on AP-10K while using far fewer parameters than ViTPose. Ablation studies on MPII examine block counts, channel widths, the stem, and the CMM.","tokens_in":15387,"tokens_out":9260,"duration_ms":72533,"significance":"If the headline comparisons were properly matched, this would be a useful baseline: the model is simple, the parameter counts are low, and the ablation results in Section 4.3 are internally consistent and support the value of the stem and CMM. The claim of being the first Mamba-based 2D keypoint detector is plausible and gives the community a concrete reference architecture. However, the paper ships no code, and the load-bearing speed and SOTA claims are currently supported only by mismatched experimental protocols. The contribution is therefore promising but not yet validated at the level claimed.","major_comments":[{"comment":"The speed comparison is not protocol-matched and cannot support the headline FPS claims. MamKPD is a top-down method (Section 4.1 states that an instance detector runs first), so the reported 1492-2564 FPS appears to measure only the keypoint branch on cropped instances. In contrast, RTMO and KAPAO are one-stage methods whose quoted FPS includes person detection and keypoint prediction in a single forward pass. Furthermore, the text notes that ViTPose numbers were measured on an A100 while MamKPD was measured on an RTX 4090, and no input resolution, batch size, numerical precision, or post-processing details are given for any of the speed measurements. Because the '4x faster than ViTPose-L' and real-time advantages rest entirely on these numbers, the authors should re-benchmark all models on the same GPU with a fixed protocol, reporting either end-to-end FPS with a detector or explicitly labeled pose-branch-only FPS against matched pose-branch baselines.","section":"Section 4.2.2, Table 2"},{"comment":"The state-of-the-art claim on MPII is undermined by the dagger footnote. ViTPose-S/B, HRNet-W32, and TokenPose-T are reproduced without ImageNet pre-training, while the published comparison numbers, such as TokenPose-L, use pre-trained backbones. Since MamKPD is also trained from scratch, the 91.3 mean PCK demonstrates superiority only over randomly-initialized re-implementations, not over the published SOTA. The AP-10K comparison in Table 4 has the same issue for the marked baselines. Please use published numbers with the standard pre-training protocol, retrain all baselines under a matched protocol, or explicitly restrict the SOTA claim to the from-scratch setting.","section":"Section 4.2.1, Tables 3 and 4"},{"comment":"Several experimental details needed to interpret the efficiency claims are missing: the input resolution used for training and testing, the batch size, the number of warm-up iterations and repeated runs for FPS, whether FP16 or FP32 was used, and whether a flip test or other post-processing was applied at evaluation. The GFLOPs values in Table 1 depend on input resolution, so without this information the parameter-efficiency and speed comparisons cannot be reproduced or verified. These details should be added to the implementation-details paragraph.","section":"Section 4.1, Tables 1 and 2"}],"minor_comments":[{"comment":"HRNet-W48 is listed with 28.5M parameters, which is inconsistent with Table 2 (63.6M) and with the cited paper; please correct the table.","section":"Section 4.2.1, Table 4"},{"comment":"NVIDIA GTX 4090 should read NVIDIA RTX 4090.","section":"Abstract and Section 4.2.2"},{"comment":"The word 'conmmunication' is a typo for 'communication'.","section":"Introduction, Section 3.2"},{"comment":"The Dite-HRNet-18 row is duplicated with different scores; remove the duplicate or label the intended variants.","section":"Table 3"},{"comment":"The notation should clarify whether the SS2D operation is applied once per stage or repeated N_i times per stage, since Figure 2(c) and Table 1 indicate N_i blocks.","section":"Equation (2) and Figure 2(c)"},{"comment":"The claim that the conventional Mamba module has limited information interaction between patches should be reconciled with the fact that SS2D performs four-directional global scans; otherwise the motivation for CMM is confusing.","section":"Section 3.2"},{"comment":"The abstract says MamKPD saves '85% of the parameters compared to ViTPose' without specifying ViTPose-B; please name the exact comparison model, since ViTPose-L would give a different percentage.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The stress-test critique is on point. The architecture is worth publishing as a baseline, but the current draft overclaims because of unmatched speed and pretraining protocols. I would accept a revised version that provides same-GPU, same-protocol speed measurements and matched MPII baselines. The novelty claim of being the first Mamba-based 2D keypoint detector is plausible but should be checked against recent literature before final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is the first Mamba-based 2D keypoint detector, and the core idea is clean: keep the encoder simple, add a small context module to fix Mamba's weak inter-patch interaction, and use a standard decoder. The results on COCO and MPII are new measurements, and the parameter efficiency is genuinely striking—MamKPD-L at 12.4M parameters beats ViTPose-B (86M) by 1.5 AP on COCO while using no pretraining. That is worth taking seriously.\n\nThe internal ablations are coherent: CMM adds ~3.2 points over the pure Mamba baseline, and the stem adds another ~1.3. The design choices are reasonable, and the paper is honest about the A100/4090 speed discrepancy in Section 4.2.2.\n\nThe soft spots are all about the empirical comparisons, and they are not minor. The headline 1492 FPS is a pose-branch-only number for a top-down method, while RTMO and KAPAO are one-stage methods reporting end-to-end FPS that includes detection. Comparing those is like comparing a car's engine-only horsepower to a full-car drag time. The ViTPose speed numbers come from an A100, so even among top-down methods the speed column is not matched. Input resolution, batch size, FP16/FP32, flip test, and detector time are not reported. The MPII table reproduces ViTPose and HRNet without ImageNet pretraining while keeping other baselines pretrained, which makes the 'state-of-the-art on MPII' claim fragile. None of this invalidates the architecture, but it does mean the paper's central selling point—dramatically higher speed at equal accuracy—is not yet supported.\n\nThe 4-stage ablation losing 6.5 points is odd but likely overfitting, and the authors say so; I would not make a big deal of it.\n\nWho should read this: anyone building efficient pose estimators or applying Mamba to dense prediction. It is a solid baseline paper, not a scientific breakthrough. The math is trivial, but the measurements are real. What it needs before the strong claims can stand: matched speed benchmarks on the same GPU with the same protocol, one-stage full-pipeline numbers, pretraining-aware comparisons, and code.\n\nI would send it to peer review. The architecture is novel enough, and the efficiency results, if reproduced, would be useful to the community. But the reviewer should demand a fair benchmark and code release before accepting the speed and SOTA claims.","headline":"A genuinely first Mamba-based 2D pose estimator with real parameter efficiency, but the headline speed claim rests on unmatched comparisons and needs fairer benchmarks and code before it can be believed.","tokens_in":15922,"tokens_out":2986,"would_cite":true,"duration_ms":25871,"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":"MamKPD, a Mamba pose model, reaches 77.3% AP at 1492 FPS on COCO.","keywords":["2D keypoint detection","human pose estimation","animal pose estimation","Mamba","state space models","real-time inference","contextual modeling module","heatmap regression"],"falsifier":"Retrain ViTPose-B and RTMO-L on COCO train2017 without ImageNet pre-training, benchmark all models on the same RTX 4090 with the same input resolution, batch size, and precision, and compare AP and FPS; if ViTPose-B or RTMO-L matches or beats MamKPD-L on both axes, the claimed Mamba advantage does not hold under matched conditions.","tokens_in":14927,"feed_emoji":"🦴","tokens_out":6097,"duration_ms":49089,"temperature":0.7,"pith_summary":"This paper proposes MamKPD, a heatmap-based, top-down 2D keypoint detection network built on a Mamba state-space backbone instead of a CNN or Transformer. The authors claim it is the first Mamba architecture designed for 2D keypoint detection, and that it removes a known Mamba weakness by adding a lightweight contextual modeling module that lets image patches exchange information. On COCO val2017, the large variant reaches 77.3% average precision at 1492 FPS on a single RTX 4090 GPU, with roughly 4% of the parameters of the comparable ViTPose-L model. The paper therefore argues that Mamba can replace both CNN and Transformer backbones in a pose estimator that is simultaneously accurate and very fast. If true, the result matters because real-time pose estimation would no longer need heavyweight attention or deep convolutional stacks.","feed_headline":"Mamba pose model hits 77.3% AP at 1492 FPS","feed_subtitle":"Paper claims this Mamba-based model matches transformer accuracy on COCO with an 85% smaller network.","key_machinery":"The load-bearing mechanism is the contextual modeling module (CMM) used inside every stage. CMM is a lightweight two-part block: a depth-wise convolution along the patch dimension, followed by a patch-embedding layer and normalization, models dependencies between patches and expands the receptive field; a linear layer plus a 1x1 depth-wise convolution then distills pose cues within each patch. This is interleaved with the SS2D block, which scans the feature map in four directions to aggregate global patch information. The combination is what lets Mamba see local anatomical context before global scanning, and the paper's ablations attribute a 3.2-point mean-PCK gain on MPII to the CMM.","core_discovery":"The paper's central claim is that a carefully staged Mamba encoder can deliver accurate 2D keypoint detection without the parameter cost of transformers. Called MamKPD, the network uses a CNN stem, three encoder stages, and a two-deconvolution decoder; each stage pairs a contextual modeling module with 2D selective-scan SS2D blocks. The contextual modeling module uses depth-wise convolutions over the patch grid to capture inter-patch dependencies and a linear layer with a 1x1 depth-wise convolution to distill pose cues inside each patch. The paper reports 77.3% AP on COCO val2017 for MamKPD-L at 1492 FPS, 91.3 mean PCK on MPII, and competitive AP-10K animal pose results, while saving roughly 85% of parameters compared to ViTPose. The authors present the result as evidence that Mamba is a viable backbone class for pose estimation, not just a sequential modeling tool.","pith_inferences":["The paper does not explore whether CMM helps other Mamba vision backbones beyond MamKPD; if it does, the same module could be reused in segmentation or detection heads.","The reported FPS numbers come from an RTX 4090 with no matched hardware benchmark against all baselines, so a head-to-head evaluation with identical GPUs, input resolutions, batch sizes, and precision would settle the speed comparison.","The parameter savings suggest potential deployment on embedded devices, but the paper only measures a desktop GPU, so edge-latency remains untested.","The staged Mamba encoder could likely be adapted to single-stage or bottom-up multi-person pose estimation, but those formulations are not explored here."],"forward_implications":["A Mamba backbone can serve as the core of a top-down 2D pose estimator, achieving transformer-level COCO AP while using a small fraction of the parameters.","Real-time pose estimation can run at more than 1400 FPS on a single high-end consumer GPU, according to the paper's measurements.","Adding local patch-context modeling fixes the main weakness of naive Mamba for images, and the paper shows the gain on MPII.","The design transfers beyond humans: MamKPD also reports competitive animal keypoint results on AP-10K.","Because the training uses only MSE heatmap loss with no extra supervision, the architecture is a simple baseline that later Mamba pose methods can build on."],"supporting_citations":[{"why":"Supplies the selective state-space model architecture that MamKPD uses for global patch modeling.","marker":"[11]"},{"why":"Provides the four-direction 2D selective-scan (SS2D) block used in each MamKPD stage.","marker":"[29]"},{"why":"Defines the encoder-decoder structure, heatmap MSE loss, and simple decoder that MamKPD adopts.","marker":"[50]"},{"why":"Serves as the main transformer baseline for accuracy, parameter, and speed comparisons.","marker":"[52]"},{"why":"Provides the high-resolution CNN baseline and the training and augmentation protocol followed by the paper.","marker":"[42]"},{"why":"Defines the COCO dataset and AP metrics used for the main accuracy and speed evaluation.","marker":"[26]"},{"why":"Defines the MPII dataset and PCK metric used for keypoint-level evaluation.","marker":"[1]"},{"why":"Defines the AP-10K animal pose benchmark used for cross-domain evaluation.","marker":"[57]"},{"why":"Provides the RTMO real-time one-stage baseline that MamKPD compares against for speed and accuracy.","marker":"[31]"}],"fun_headline_variants":["Mamba-based pose detector achieves 77.3% AP, 1492 FPS","First Mamba-based real-time keypoint detector: 77.3% AP on COCO","MamKPD: Mamba baseline for 2D keypoints, 85% fewer params","Real-time pose estimation with Mamba: 77.3% AP, 1492 FPS","Mamba matches transformer pose accuracy at 1492 FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed efficiency-and-accuracy advantage assumes all comparison numbers were measured under the same protocol, because MamKPD's speed was measured on an RTX 4090 while some transformer baselines report A100 numbers, and several reproduced baselines were trained without ImageNet pre-training while others used it.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-based pose detector achieves 77.3% AP, 1492 FPS","First Mamba-based real-time keypoint detector: 77.3% AP on COCO","MamKPD: Mamba baseline for 2D keypoints, 85% fewer params","Real-time pose estimation with Mamba: 77.3% AP, 1492 FPS","Mamba matches transformer pose accuracy at 1492 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000412,"raw_usage":{"total_tokens":2144,"prompt_tokens":969,"completion_tokens":1175,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":1060}},"tokens_in":585,"tokens_out":1175,"duration_ms":7294,"temperature":1.0,"reasoning_tokens":1060,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:23:17.640312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain ViTPose-B and RTMO-L on COCO train2017 without ImageNet pre-training, benchmark all models on the same RTX 4090 with the same input resolution, batch size, and precision, and compare AP and FPS; if ViTPose-B or RTMO-L matches or beats MamKPD-L on both axes, the claimed Mamba advantage does not hold under matched conditions.","supporting_citations":[{"cited_title":"Simple baselines for human pose estimation and tracking","cited_arxiv_id":null,"evidence_quote":"Defines the encoder-decoder structure, heatmap MSE loss, and simple decoder that MamKPD adopts."},{"cited_title":"Vit- pose: Simple vision transformer baselines for human pose estimation","cited_arxiv_id":null,"evidence_quote":"Serves as the main transformer baseline for accuracy, parameter, and speed comparisons."},{"cited_title":"Deep high-resolution representation learning for human pose esti- mation","cited_arxiv_id":null,"evidence_quote":"Provides the high-resolution CNN baseline and the training and augmentation protocol followed by the paper."},{"cited_title":"Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C","cited_arxiv_id":null,"evidence_quote":"Defines the COCO dataset and AP metrics used for the main accuracy and speed evaluation."},{"cited_title":"Gehler, and Bernt Schiele","cited_arxiv_id":null,"evidence_quote":"Defines the MPII dataset and PCK metric used for keypoint-level evaluation."},{"cited_title":"AP-10K: A benchmark for animal pose esti- mation in the wild","cited_arxiv_id":null,"evidence_quote":"Defines the AP-10K animal pose benchmark used for cross-domain evaluation."},{"cited_title":"RTMO: towards high-performance one- stage real-time multi-person pose estimation","cited_arxiv_id":null,"evidence_quote":"Provides the RTMO real-time one-stage baseline that MamKPD compares against for speed and accuracy."}],"review_version":1}