{"id":"14559075-9b1f-476f-8bb7-4233ebc3ccbb","arxiv_id":"2506.14596","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PoseGRAF fuses joint and bone-direction features through cross-attention and dynamic fusion, reporting marginal state-of-the-art MPJPE improvements on Human3.6M and MPI-INF-3DHP.","lead":"This paper presents PoseGRAF, a neural network that lifts 2D body keypoints to 3D poses by combining joint and bone graphs with cross-attention and dynamic fusion. It reports small accuracy gains over prior methods on two standard 3D pose benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MPI-INF-3DHP evaluation lacks a defined skeleton mapping, so the cross-dataset SOTA claim is unverified.","rationale":"I read the paper's central claim as that PoseGRAF exceeds SOTA on Human3.6M and MPI-INF-3DHP. The Human3.6M results are internally consistent: Table 1 shows 48.1 mm MPJPE vs. GraphMLP's 49.2 mm (a 1.1 mm gap), and the '10.6 mm' overstatement in Sec. 4.3 is a numerical error that does not change the reported ordering. The more fundamental gap is the MPI-INF-3DHP protocol: the text says the model is applied 'directly' after training on Human3.6M, but no joint/skeleton mapping is given. Because the architecture includes dataset-specific graphs (G_J, bone graph with M bones, BFS), the model is not architecture-agnostic to skeleton changes; the authors must specify how N and M are set and how PCK is computed. Prior works use varied protocols (e.g., 14-joint vs. 17-joint), so without this the 88.3 PCK is not comparable. The concrete test of inspecting the released code would settle it. The reader also flagged missing μ and w; I agree those are important reproducibility gaps, but they are secondary to the evaluation protocol issue. My verdict remains CONDITIONAL: acceptable after the authors clarify the cross-dataset protocol and provide the missing hyperparameter values. This does not change the reader's verdict.","tokens_in":13825,"tokens_out":7734,"duration_ms":78012,"concrete_test":"Check the released code's MPI-INF-3DHP evaluation script. Verify the joint mapping (model output to ground truth) and the graph construction (G_J, number of bones M, adjacency) applied to MPI-INF-3DHP. Reproduce Table 3 using the exact protocol of GraphMLP or DGFormer (e.g., 17-joint COCO mapping); if the 88.3 PCK / 54.8 AUC numbers hold under that protocol, the concern is resolved. If the code maps joints differently, recompute PCK/AUC under the standard protocol and see whether the SOTA margin persists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 3's SOTA claim on MPI-INF-3DHP rests on applying a Human3.6M-trained model 'directly' (Sec. 4.3), yet the paper never defines the joint correspondence or graph adjacency between the two datasets. Human3.6M (17 COCO joints via CPN) and MPI-INF-3DHP have different ground-truth skeletons; the Dynamic-Fusion BFS (Eq. 9) and the top-μ selection depend on G_J, so the architecture can only transfer if the same skeleton is imposed. Without specifying how N, M, G_J, and the PCK evaluation mapping are set, the reported 88.3 PCK cannot be compared to prior results. This is the more load-bearing issue because it directly affects the second pillar of the 'exceeds state-of-the-art' claim; the missing hyperparameters w (Eq. 11) and μ (Algorithm 1) are secondary reproducibility gaps.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes PoseGRAF, a 2D-to-3D monocular human pose estimation framework that combines joint and bone graph convolutions, a cross-attention module, an attention-based dynamic fusion module, and a transformer encoder with relative-distance attention scaling. The authors report state-of-the-art MPJPE (48.1 mm) and P-MPJPE (38.3 mm) on Human3.6M with CPN-detected 2D poses, and PCK (88.3%) and AUC (54.8%) on MPI-INF-3DHP when a Human3.6M-trained model is applied directly. Ablations on Human3.6M ground-truth 2D poses attribute gains to the bone-direction GCN, cross-attention, and dynamic fusion. The code is stated to be publicly available.","tokens_in":13974,"tokens_out":5051,"duration_ms":43130,"significance":"If the reported numbers are reproducible, PoseGRAF offers a modest but useful improvement over the closest competitor GraphMLP on both benchmarks, and the ablation study provides evidence that the geometric bone-direction modeling contributes positively. The release of code is a practical strength. However, the margins over prior work are small (1.1 mm MPJPE, 0.3 mm P-MPJPE, 1.3 PCK, 0.5 AUC), the paper overstates the GraFormer comparison, and key hyperparameters and the cross-dataset protocol are underspecified. The central architectural idea is defensible, but the current presentation does not yet provide enough detail to verify the strongest claims.","major_comments":[{"comment":"The claim in Section 4.3 that PoseGRAF achieves 'a reduction of 10.6 mm in MPJPE over GraFormer' is arithmetically inconsistent with Table 1, which lists GraFormer at 51.8 mm and PoseGRAF at 48.1 mm, a gap of 3.7 mm. The text should either be corrected to 3.7 mm or the intended comparison should be stated explicitly.","section":"4.3, Table 1"},{"comment":"The MPI-INF-3DHP evaluation is described as applying the Human3.6M-trained model 'directly' (Section 4.3), but the manuscript does not specify the joint correspondence, the joint count N, the skeletal adjacency matrix G_J, or the bone-node construction used for MPI-INF-3DHP. Since Eqs. (1)-(10) and Algorithm 1 depend on G_J and the bone graph, the reported 88.3 PCK is not verifiable against prior comparisons unless these mapping details are provided.","section":"4.3, Table 3"},{"comment":"The hyperparameters w in Eq. (11) and the top-μ selection in Algorithm 1 are load-bearing but never specified. Section 3.5 states only that 'we set w to facilitate information exchange...' without a value or tuning procedure, and Section 3.4 describes μ as both a 'learnable threshold parameter' and the count of selected top-μ joint features, while Algorithm 1's Top_Indices call does not define μ. Please provide explicit values, ranges, or learning rules for both.","section":"3.4, Algorithm 1 and 3.5, Eq. (11)"},{"comment":"The reported improvements over GraphMLP are small (1.1 mm MPJPE, 0.3 mm P-MPJPE, 1.3 PCK, 0.5 AUC), yet no error bars, multiple-seed variation, or significance tests are reported. Without such information, the claim of 'exceeding state-of-the-art' is not robust to run-to-run variability; at minimum, the authors should report standard deviations across seeds for the main tables.","section":"Tables 1-3"}],"minor_comments":[{"comment":"The denominator of the angle formula should be ‖x_B^p‖‖x_B^q‖, not ‖x_B^p‖‖x_B^p‖.","section":"Eq. (3)"},{"comment":"The caption of Table 1 uses '&' and '*' markers, while Table 2's caption uses '*'; ensure the symbols are defined consistently across tables and that the rows actually carry the markers.","section":"Tables 1-2"},{"comment":"The text says static fusion 'increases MPJPE by 4.1 mm' compared to dynamic fusion, but Table 4 does not contain the static-fusion pair needed to verify this number; please add the missing ablation row or clarify the reference point.","section":"Section 4.4, Table 4"},{"comment":"Implementation details list architecture and training hyperparameters but omit the values of w and μ; these should be included alongside the other settings.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal. My main concerns are the overstatement in Section 4.3, the unspecified cross-dataset skeleton mapping, and the missing hyperparameter values; these are fixable in revision. I did not attempt to run the provided code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: PoseGRAF is a competent, incremental 2D-to-3D pose lifting paper that reports small SOTA gains on two standard benchmarks. The architecture is a sensible combination of known components — bone-angle weighted graphs, cross-attention between joint and bone features, and a BFS-based dynamic fusion module — and the ablation study supports each addition. The main tables are internally consistent, and the code is promised. If you work on this benchmark treadmill, the 1 mm gains over GraphMLP are plausible and worth a look.\n\nWhere it gets soft: the MPI-INF-3DHP result rests on applying the Human3.6M-trained model 'directly' without specifying the skeleton mapping. Human3.6M and MPI-INF-3DHP have different ground-truth joint sets, and the dynamic fusion BFS depends on G_J, so the reported PCK/AUC can't be compared to prior work unless the joint correspondence is defined. That is the load-bearing issue for the second SOTA claim. Separately, the text says PoseGRAF reduces MPJPE over GraFormer by 10.6 mm, but the actual gap in Table 1 is 3.7 mm — a clear overstatement. Also missing: the values of w (Eq. 11) and μ (Algorithm 1), which are needed to reproduce the dynamic fusion and attention scaling. There are no error bars or significance tests, so the 1.1 mm edge over GraphMLP could be noise. Minor typos (Eq. 3 denominator) don't help.\n\nOverall: the central idea is reasonable, the ablations are informative, and the method is neither circular nor a fitted constant. It's a modest contribution that should go through peer review, but the authors need to fix the cross-dataset evaluation, correct the numeric claim, and specify the missing hyperparameters. I'd bring it to a reading group for the architecture discussion, but I wouldn't cite it myself this year.","headline":"Competent incremental pose lifting paper with plausible small gains; the cross-dataset SOTA claim is under-specified and the GraFormer comparison is misquoted.","tokens_in":14522,"tokens_out":2588,"would_cite":false,"duration_ms":25252,"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":"PoseGRAF claims that adding bone-direction and joint-angle graphs to a 2D-to-3D lifting network improves monocular 3D pose accuracy, reporting top MPJPE, P-MPJPE, and PCK results on Human3.6M and MPI-INF-3DHP.","keywords":["3D human pose estimation","monocular 2D-to-3D lifting","graph convolutional networks","bone direction","joint angles","dynamic feature fusion","cross-attention","Human3.6M"],"falsifier":"On Human3.6M ground-truth 2D input, sweep the unspecified parameters—$w$ from 0 to 5 and $\\mu$ from 1 to 17—and re-run the MPI-INF-3DHP transfer with the Human3.6M joint ordering explicitly remapped; if MPJPE or PCK moves by more than the reported margins over GraphMLP, the headline gains are not robust to these undisclosed choices.","tokens_in":13599,"feed_emoji":"🦴","tokens_out":6649,"duration_ms":62232,"temperature":0.7,"pith_summary":"The paper tries to establish that monocular 3D human pose estimation from 2D keypoints improves when the skeleton is represented not only by joint positions but also by bone directions and the angles between adjacent bones. It builds a dual-graph architecture, a cross-attention mechanism between joint and bone features, and an adaptive fusion stage that selects key joints and reconstructs global features by traversing the skeleton graph. On Human3.6M, the method reports the best MPJPE and P-MPJPE with CPN-detected 2D poses (48.1 mm and 38.3 mm), and on MPI-INF-3DHP it reports the best PCK (88.3) among the compared methods. The claim matters because implausible poses under occlusion and fast motion are a known failure mode of lifting methods, and the paper's design targets that failure directly.","feed_headline":"Bone-angle graphs cut monocular 3D pose error to 48.1 mm","feed_subtitle":"Dual joint-and-bone graph with adaptive fusion tops Human3.6M and MPI-INF-3DHP tables.","key_machinery":"The load-bearing component is a Bone Direction GCN that builds two bone graphs: a directed weighted graph whose edge weights are the angles between adjacent bone direction vectors, computed via $\\arccos$ of their normalized inner product, and a directed unweighted graph of bone connectivity. Beside it, a Joint GCN processes the usual joint-position graph. A Cross-Attention module mixes joint and bone features, and a Dynamic Fusion module selects the top-$\\mu$ joints by attention score, then propagates from those seeds across the joint graph with a breadth-first search while injecting bone-direction features, and adds the original joint features back. An improved Transformer encoder rescales attention weights with a relative distance matrix derived from the skeleton topology and is joined to the rest in a residual manner.","core_discovery":"The central claim, stated on the paper's own terms, is that explicitly modeling bone directions and the angles between adjacent bones as weighted edges in a graph, then adaptively fusing those geometric features with joint-position features through cross-attention and a top-$\\mu$ breadth-first reconstruction, lets a 2D-to-3D lifting network exceed the prior state of the art on both Human3.6M and MPI-INF-3DHP. The reported numbers are 48.1 mm MPJPE and 38.3 mm P-MPJPE with CPN-detected 2D poses on Human3.6M, 32.1 mm MPJPE with ground-truth 2D poses, and 88.3 PCK with 54.8 AUC on MPI-INF-3DHP using a model trained on Human3.6M with no stated adaptation.","pith_inferences":["A testable extension the paper does not run: sweep the unspecified parameters $w$ and $\\mu$ and report the error surface; if the gains over GraphMLP vanish outside a narrow band, the adaptive fusion is not the true source of improvement.","Because every module assumes a fixed skeleton graph, transferring to a dataset with a different joint convention would require remapping joint indices and recomputing the bone graphs; the paper does not discuss this implicit cost.","The comparisons are single-frame; inserting the same geometric fusion into a temporal encoder, as other lifted pose models do, is a natural next step that the paper leaves open."],"forward_implications":["With CPN-detected 2D input on Human3.6M, PoseGRAF reports 48.1 mm MPJPE and 38.3 mm P-MPJPE, the best averages in the comparison tables.","With ground-truth 2D input, it reports 32.1 mm MPJPE and 25.0 mm P-MPJPE, a 7.2% reduction over its own Transformer-plus-Joint-GCN baseline.","A Human3.6M-trained model reports 88.3% PCK and 54.8% AUC on MPI-INF-3DHP with no stated adaptation to the second dataset.","Ablations show that adding the Bone GCN and Cross-Attention lowers error monotonically and that the full dynamic-fusion configuration is the best performer, supporting the design's core premise."],"supporting_citations":[{"why":"Supplies the Human3.6M benchmark and the training/evaluation protocol all reported comparisons use.","marker":"[13]"},{"why":"Supplies the MPI-INF-3DHP benchmark and the PCK/AUC protocol for the cross-dataset evaluation.","marker":"[14]"},{"why":"Provides the CPN 2D keypoint detector whose outputs feed every reported 2D-to-3D input.","marker":"[30]"},{"why":"GraphMLP is the closest prior method in the CPN-input comparison, against which the claimed 1.1 mm MPJPE gain is measured.","marker":"[40]"},{"why":"DGFormer supplies the strongest graph-transformer baseline under ground-truth 2D input, edged out by 0.3 mm MPJPE.","marker":"[8]"},{"why":"Cited as the inspiration for the attention-based dynamic fusion and top-$\\mu$ feature selection.","marker":"[32]"},{"why":"Cited for the relative-distance matrix used to rescale attention weights inside the improved Transformer encoder.","marker":"[34]"}],"fun_headline_variants":["Bone graphs beat joint-only 3D pose lifters","PoseGRAF fuses bone angles to top 3D pose benchmarks","Joint-bone cross-attention boosts 3D pose accuracy","Geometric graphs sharpen 3D pose estimation","Bone-angle graph fusion cuts 3D pose error to 48.1mm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on the unstated assumptions that the two datasets share an identical joint set, ordering, and skeletal connectivity used to build the joint graph, bone graphs, and BFS, and that the distance-scaling weight $w$ and top-$\\mu$ selection are set appropriately.","fun_headline_variants_meta":{"raw":{"variants":["Bone graphs beat joint-only 3D pose lifters","PoseGRAF fuses bone angles to top 3D pose benchmarks","Joint-bone cross-attention boosts 3D pose accuracy","Geometric graphs sharpen 3D pose estimation","Bone-angle graph fusion cuts 3D pose error to 48.1mm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000832,"raw_usage":{"total_tokens":3618,"prompt_tokens":919,"completion_tokens":2699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":2608}},"tokens_in":535,"tokens_out":2699,"duration_ms":19159,"temperature":1.0,"reasoning_tokens":2608,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:50:17.502053+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On Human3.6M ground-truth 2D input, sweep the unspecified parameters—$w$ from 0 to 5 and $\\mu$ from 1 to 17—and re-run the MPI-INF-3DHP transfer with the Human3.6M joint ordering explicitly remapped; if MPJPE or PCK moves by more than the reported margins over GraphMLP, the headline gains are not robust to these undisclosed choices.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Human3.6M benchmark and the training/evaluation protocol all reported comparisons use."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MPI-INF-3DHP benchmark and the PCK/AUC protocol for the cross-dataset evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CPN 2D keypoint detector whose outputs feed every reported 2D-to-3D input."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GraphMLP is the closest prior method in the CPN-input comparison, against which the claimed 1.1 mm MPJPE gain is measured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DGFormer supplies the strongest graph-transformer baseline under ground-truth 2D input, edged out by 0.3 mm MPJPE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as the inspiration for the attention-based dynamic fusion and top-$\\mu$ feature selection."},{"cited_title":"J., Jacobs, D","cited_arxiv_id":null,"evidence_quote":"Cited for the relative-distance matrix used to rescale attention weights inside the improved Transformer encoder."}],"review_version":2}