{"id":"adf09eaa-f005-4f84-ab4c-d4a0ae6dc2d6","arxiv_id":"1908.10357","paper_version":3,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A bottom-up pose estimator built by adding a deconvolution-based high-resolution feature pyramid to HRNet, trained with multi-resolution supervision and tested with heatmap aggregation, reports state-of-the-art COCO and CrowdPose results.","lead":"This paper presents HigherHRNet, a neural network that estimates the poses of many people in one image by predicting keypoints at multiple image resolutions and combining the results. It reports the best accuracy of bottom-up pose estimators on COCO and CrowdPose benchmarks, especially for medium-sized and crowded people.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark claim is supported; the scale-aware attribution is the softest point, but it does not undermine the empirical core.","rationale":"The reader identified the same weakest assumption: the component-wise gains and the choice of deconvolution count and training size are selected on COCO val, and the ablation tables are single runs, so the scale-aware interpretation is less rigorously pinned down than the headline numbers. My stress-test agrees with that reading. I did not find an internal inconsistency or a defect in the benchmark comparisons: the paper is transparent that the 70.5 AP result uses multi-scale test, the single-scale result is also reported, and the comparison tables include prior bottom-up and top-down methods with the same evaluation protocol. The central claim as stated in Section 4.1 is an empirical state-of-the-art claim, and the evidence (public benchmark numbers, ablations, code link) is appropriate for an ACCEPT verdict. The scale-aware explanation is a real but non-fatal soft spot: the deltas in Table 4 are small and unreplicated, and the paper explicitly declines to assign different scales to different pyramid levels, so the mechanism is more of a plausible interpretation than a demonstrated causal effect. Because the verdict is based on the benchmark result rather than on the mechanism, no verdict change is needed. The concrete test I propose—multi-seed replication of Table 4 and the deconvolution-count comparison—would settle whether the attribution concern lands, and if it does, it would affect only the explanatory framing, not the empirical contribution.","tokens_in":13339,"tokens_out":4450,"duration_ms":45517,"concrete_test":"Use the released code to rerun the five configurations in Table 4 and the one-versus-two deconvolution comparison in Table 3 on COCO val with at least three random seeds per configuration, reporting mean and standard deviation of AP, APM, and APL. If the component deltas (e.g., feature concatenation +0.3 AP, heatmap aggregation +0.6 AP) are within one standard deviation or change sign across seeds, the scale-aware attribution should be softened in any future revision, although the headline test-dev comparison would be unaffected. A secondary check is to evaluate the released W48 model on COCO test-dev (or a held-out split) to independently confirm the 70.5 AP figure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is an empirical benchmark result (COCO test-dev 70.5 AP, CrowdPose 67.6 AP), not a derivation, and the reported evidence—comparisons to prior bottom-up methods, component ablations on COCO val, and a public code release—supports that claim as far as a paper review can. The load-bearing risk I can identify is therefore not the headline number but the scale-aware explanation attached to it: the attribution of the gains to the specific proposed mechanisms rests on single-run ablation deltas (Table 4: 66.0 to 66.3 to 66.9 to 67.1 AP from adding feature concatenation, heatmap aggregation, and residual blocks) and on model choices (one versus two deconvolution modules, training size 512/640/768, Tables 3 and 5) that were selected on the same COCO val set. These deltas are within the range where seed-level or optimization noise could change the ordering; for example, feature concatenation improves AP by only 0.3. Thus the causal story 'scale-aware feature pyramid helps small persons' is less firmly established than the benchmark claim itself. This does not undermine the benchmark comparison, but it is the softest load-bearing point if the paper is read as demonstrating scale-aware representation learning rather than just reporting state-of-the-art results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces HigherHRNet, a bottom-up multi-person 2D pose estimation method that augments HRNet with a high-resolution feature pyramid built from transposed-convolution deconvolution modules, trained with multi-resolution supervision and combined at inference by multi-resolution heatmap aggregation. The authors report 70.5 AP on COCO 2017 test-dev without refinement, exceeding previous bottom-up methods, and 67.6 AP on CrowdPose test, exceeding all compared top-down and bottom-up methods. Ablations on COCO val (Tables 3-6) decompose the gains from the deconvolution module, feature concatenation, heatmap aggregation, residual blocks, training resolution, and backbone width.","tokens_in":13567,"tokens_out":10416,"duration_ms":104787,"significance":"If the reported results are correct, the paper is a clear advance: it establishes a simple, efficient high-resolution feature-pyramid design as the state of the art in bottom-up pose estimation and provides a strong bottom-up result on CrowdPose. The paper's strengths are that it follows standard COCO/CrowdPose evaluation protocols, reports single-model numbers, compares against a wide set of prior methods, includes component-level ablations, and ships code and pretrained models. The conceptual claim that the gains are due to scale-aware multi-resolution representation is plausible and supported by the monotonic trend in the ablations, although the component-level evidence is less tight than the benchmark result itself; I agree with the stress-test assessment that the scale-aware attribution is the softest point, but it does not undermine the empirical core.","major_comments":[],"minor_comments":[{"comment":"The component ablations are single-run, and several design choices (number of deconvolution modules, training image size) are selected on the same COCO val set used for the ablations; because the individual deltas are small (e.g., feature concatenation improves AP by only 0.3 in Table 4), please state this limitation explicitly and either provide multi-seed variance or temper the causal language so that the cumulative monotonic trend, rather than any single delta, is presented as the evidence for the design.","section":"4.2, Tables 4-6"},{"comment":"The phrase \"especially for small person\" overstates what can be measured, since the COCO keypoint task has no small-person annotations; the supporting numbers are for medium persons (APM), so please rephrase to \"smaller/medium persons\" or otherwise clarify the proxy used.","section":"Abstract and Section 4.1"},{"comment":"The statement \"We believe the model has the ability to automatically focus on specific scales\" is an assertion rather than a demonstrated property; please either add evidence (for example, per-scale heatmap visualizations) or mark it clearly as a conjecture.","section":"Section 3.4"},{"comment":"The sentence \"Comparing method (d) and (e), the gain of heatmap aggregation comes from large person\" should compare methods (c) and (d); as written, it attributes to heatmap aggregation the step that is actually the addition of residual blocks.","section":"Table 4 text"},{"comment":"Please specify the training and testing image size used for the backbone comparison in Table 6 (the text does not state that it uses the 640x640 setting), and add a footnote clarifying the testing protocol for each ablation table.","section":"Tables 5 and 6"},{"comment":"The phrase \"without any post processing\" is qualified by the use of flip testing and multi-scale test; please clarify whether multi-scale test is considered test-time augmentation rather than post-processing, so that readers do not infer that no test-time operations are used.","section":"Abstract and Table 1"},{"comment":"The statement \"All training parameters follow COCO exactly\" conflicts with the later statement that a crop size of 640x640 is used; please state that the crop size is the only intentional difference.","section":"Section 4.3"},{"comment":"There are several small typos: \"netork\" in Section 2, \"concatennation\" in Figure 3's caption, \"muilt\" in Section 5, \"MSR\" for \"MRS\" in Table 4's caption, and \"headmaps\" in Section 3.1; these should be corrected.","section":"Throughout"}],"recommendation":"minor_revision","confidential_remarks":"This is a strong systems paper whose empirical benchmark claims are well supported by standard protocols and external test sets. The main risk is the gap between the headline scale-aware attribution and the relatively small single-run component deltas; I would not block acceptance on this, but I recommend that the authors be asked to either add a small amount of supporting analysis or soften the causal language."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the headline numbers are the main event: 70.5 AP on COCO test-dev and 67.6 on CrowdPose test, both single-model and without refinement. Those are credible SOTA results for bottom-up pose estimation, supported by standard benchmarks, sensible ablations, and a public code release. The empirical core is in good shape. Second, do not mistake the component analysis for equally solid evidence. The causal story about scale-aware representation learning is plausible but rests on small single-run deltas on the same validation set used to pick hyperparameters. That does not invalidate the AP numbers, but it means the 'scale-aware' descriptor is a reasonable interpretation, not a proven mechanism.\n\nWhat is actually new: taking HRNet's high-resolution representation and adding a deconvolution-based feature pyramid with multi-resolution supervision and heatmap aggregation is clean and well-motivated. The gains are real: on COCO val, one deconvolution module lifts AP from 64.4 to 66.9, with most of the gain on medium persons. The CrowdPose results are genuinely interesting because bottom-up methods have a natural advantage in crowds, and this paper quantifies that.\n\nSoft spots, in order. (1) The ablations are single runs, and some deltas are tiny: feature concatenation adds 0.3 AP, heatmap aggregation 0.6. Those could shuffle with a different seed. (2) The number of deconvolution modules and the training size are selected on COCO val, so the component claims are partly fitted to the same data they are evaluated on. This is standard practice but it weakens the attribution. (3) 'Without any post processing' is carefully worded: multi-scale test and flip test are used. That is normal test-time augmentation, not refinement, but worth remembering when comparing with numbers that do not use it. (4) The 'Bottom-up HRNet' baseline in Table 1 is their own implementation, not a reported number from the HRNet paper. It looks reasonable but is not independent.\n\nNone of this changes the verdict. The central claim is benchmark performance, and the evidence supports it. The paper deserves a serious referee and would be a good reading-group pick for anyone working on pose estimation or scale-aware architectures. For peer review, accept, with the main requests being repeated runs or error bars on ablations and a more cautious causal claim in the text.","headline":"HigherHRNet is a solid empirical contribution with credible SOTA numbers, but the scale-awareness story is softer than the headline and should be read as interpretation rather than demonstrated mechanism.","tokens_in":14149,"tokens_out":2209,"would_cite":true,"duration_ms":22155,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"HigherHRNet claims that bottom-up pose estimation can match top-down accuracy by predicting heatmaps from a high-resolution feature pyramid.","keywords":["bottom-up human pose estimation","scale-aware representation","high-resolution feature pyramid","multi-resolution supervision","heatmap aggregation","deconvolution module","HRNet","CrowdPose"],"falsifier":"Run the same architecture comparisons on a held-out validation set that was not used in any design choice—for example, a new split of COCO or a different scale-heavy benchmark—and check whether one deconvolution module, multi-resolution supervision, and heatmap aggregation still improve AP; if the gains vanish or reverse, the scale-aware explanation is unsupported. A more direct check on COCO is to compare inference using only the 1/2-resolution heatmap against inference using only the 1/4-resolution heatmap: if the medium-person gain disappears, the higher-resolution feature maps themselves, not the aggregation, are the cause.","tokens_in":1705,"feed_emoji":"🦴","tokens_out":1962,"duration_ms":62660,"temperature":0.7,"pith_summary":"The paper argues that the main obstacle for bottom-up multi-person pose estimation is scale variation: small persons are poorly localized by the usual single 1/4-resolution heatmap. To fix this, HigherHRNet predicts heatmaps at two resolutions, trains each level with its own ground-truth heatmap, and averages the levels at inference. On COCO test-dev it reaches 70.5 AP without refinement or post-processing, outperforming all existing bottom-up methods, with the largest gain on medium persons. On CrowdPose test it reaches 67.6 AP, surpassing even top-down methods. A sympathetic reader cares because bottom-up methods are faster and more end-to-end, and this result suggests they no longer need to concede accuracy on small-scale or crowded scenes.","feed_headline":"70.5 AP: HigherHRNet tops bottom-up pose on COCO","feed_subtitle":"The scale-aware network also beats all top-down methods on the crowded CrowdPose set, at 67.6 AP.","key_machinery":"The central object is the high-resolution feature pyramid, which reverses the usual direction of feature pyramids: instead of starting at 1/32 and upsampling to 1/4, it starts at the backbone's 1/4 resolution and generates a 1/2-resolution level. The level is produced by a deconvolution module consisting of a 4x4 transposed convolution, BatchNorm, ReLU, and four residual blocks, taking as input the concatenation of lower-level features and predicted heatmaps. Multi-resolution supervision trains each level with same-width Gaussian targets, and multi-resolution heatmap aggregation averages all levels at inference. This machinery is what lets the network localize small-person keypoints more precisely while retaining large-person performance.","core_discovery":"The discovery, on the authors' terms, is that scale variation in bottom-up pose estimation is best addressed by learning scale-aware high-resolution heatmaps rather than by relying on a single feature resolution or image pyramids. HigherHRNet takes HRNet's highest-resolution 1/4 feature maps and adds a deconvolution module that doubles the resolution to 1/2, producing a high-resolution feature pyramid. Training uses multi-resolution supervision with the same Gaussian standard deviation at both scales, and inference aggregates all predicted heatmaps after bilinear upsampling. This yields 70.5 AP on COCO2017 test-dev without any post-processing, including 66.6 AP for medium persons, and 67.6 AP on CrowdPose test, surpassing all existing methods there.","pith_inferences":["The choice to keep the Gaussian standard deviation fixed across resolutions is an unexamined assumption; the paper itself notes a smaller standard deviation is desirable at higher resolution, so per-level standard deviations could further separate small- and large-person performance.","COCO's keypoint task has no small-person subset, so the paper's small-person claims are tested only through medium persons; a benchmark with official small-person annotations would be the direct test of the scale-variation hypothesis.","If the scale-specialization story is correct, per-level recall curves should show that the 1/2-resolution heatmap alone recovers keypoints missed by the 1/4-resolution heatmap; this direct diagnostic is not presented in the paper.","The CrowdPose result suggests that bottom-up methods are robust in crowded scenes because they do not rely on single-person crops, but the grouping tags are trained only at low resolution; higher-resolution tag training, if made to converge, could push crowded-scene accuracy further."],"forward_implications":["Bottom-up pose estimation can now reach 70.5 AP on COCO test-dev without refinement, substantially narrowing the gap to top-down methods while keeping the efficiency advantage.","Most of the improvement is concentrated in medium persons, with large-person AP unchanged or slightly better, supporting the claim that the model is genuinely scale-aware rather than simply benefiting from more computation.","On CrowdPose, a crowded-scene benchmark, HigherHRNet surpasses all top-down methods, suggesting bottom-up approaches are better suited to crowded images where person detectors often fail.","The number of deconvolution levels should track the dataset's person-scale distribution; adding a second level helped only marginally on COCO and hurt large-person performance, so the architecture is intentionally adjustable.","Larger training image sizes (640) and wider backbones (HRNet-W48) both improve results, and the method still trains and tests in a single forward pass without post-processing."],"supporting_citations":[{"why":"Supplies the HRNet backbone whose highest-resolution 1/4 path is extended into a high-resolution feature pyramid.","marker":"[38, 40]"},{"why":"Demonstrates that deconvolution can generate high-quality, high-resolution heatmaps; the deconvolution module is directly inspired by it.","marker":"[42]"},{"why":"Provides the associative embedding grouping method and the heatmap/tagmap supervision scheme that HigherHRNet inherits.","marker":"[30]"},{"why":"PersonLab is the key bottom-up baseline that relies on large input resolution and is outperformed by HigherHRNet.","marker":"[33]"},{"why":"OpenPose is the classic bottom-up baseline with refinement that HigherHRNet surpasses on COCO.","marker":"[3]"},{"why":"PifPaf is a recent bottom-up method included in the COCO comparison that HigherHRNet outperforms.","marker":"[22]"},{"why":"The COCO dataset provides the train/val/test-dev splits, the OKS metric, and the medium/large person analysis.","marker":"[27]"},{"why":"The CrowdPose dataset and its top-down SPPE baseline are used to demonstrate robustness in crowded scenes.","marker":"[24]"}],"fun_headline_variants":["HigherHRNet: scale-aware pyramid lifts COCO pose to 70.5 AP","Scale-aware HRNet tops bottom-up pose, hits 70.5 AP","HigherHRNet solves small-person poses: 70.5 AP on COCO","High-res pyramid beats post-processing in bottom-up pose","HigherHRNet: multi-resolution heatmaps win on COCO and CrowdPose"],"cache_read_input_tokens":16256,"weakest_assumption_plain":"The load-bearing premise is that the performance gaps measured on the COCO validation set are stable effects of the proposed mechanisms, not artifacts of picking the number of deconvolution modules, the training image size, and the aggregation strategy using that same validation set.","fun_headline_variants_meta":{"raw":{"variants":["HigherHRNet: scale-aware pyramid lifts COCO pose to 70.5 AP","Scale-aware HRNet tops bottom-up pose, hits 70.5 AP","HigherHRNet solves small-person poses: 70.5 AP on COCO","High-res pyramid beats post-processing in bottom-up pose","HigherHRNet: multi-resolution heatmaps win on COCO and CrowdPose"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2913,"prompt_tokens":937,"completion_tokens":1976,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1875}},"tokens_in":553,"tokens_out":1976,"duration_ms":13030,"temperature":1.0,"reasoning_tokens":1875,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:45:51.756722+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same architecture comparisons on a held-out validation set that was not used in any design choice—for example, a new split of COCO or a different scale-heavy benchmark—and check whether one deconvolution module, multi-resolution supervision, and heatmap aggregation still improve AP; if the gains vanish or reverse, the scale-aware explanation is unsupported. A more direct check on COCO is to compare inference using only the 1/2-resolution heatmap against inference using only the 1/4-resolution heatmap: if the medium-person gain disappears, the higher-resolution feature maps themselves, not the aggregation, are the cause.","supporting_citations":[{"cited_title":"Simple baselines for human pose estimation and tracking","cited_arxiv_id":null,"evidence_quote":"Demonstrates that deconvolution can generate high-quality, high-resolution heatmaps; the deconvolution module is directly inspired by it."},{"cited_title":"Associa- tive embedding: End-to-end learning for joint detection and grouping","cited_arxiv_id":null,"evidence_quote":"Provides the associative embedding grouping method and the heatmap/tagmap supervision scheme that HigherHRNet inherits."},{"cited_title":"Person- lab: Person pose estimation and instance segmentation with a part-based geometric embedding model","cited_arxiv_id":null,"evidence_quote":"PersonLab is the key bottom-up baseline that relies on large input resolution and is outperformed by HigherHRNet."},{"cited_title":"Realtime multi-person 2d pose estimation using part affinity fields","cited_arxiv_id":null,"evidence_quote":"OpenPose is the classic bottom-up baseline with refinement that HigherHRNet surpasses on COCO."},{"cited_title":"Pifpaf: Composite fields for human pose estimation","cited_arxiv_id":null,"evidence_quote":"PifPaf is a recent bottom-up method included in the COCO comparison that HigherHRNet outperforms."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"The COCO dataset provides the train/val/test-dev splits, the OKS metric, and the medium/large person analysis."},{"cited_title":"Crowdpose: Efficient crowded scenes pose estimation and a new benchmark","cited_arxiv_id":null,"evidence_quote":"The CrowdPose dataset and its top-down SPPE baseline are used to demonstrate robustness in crowded scenes."}],"review_version":1}