{"id":"eef984c5-0619-4bf1-a1ed-87c46c327d77","arxiv_id":"2411.18944","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"WTPose combines a Swin transformer backbone with a multi-scale waterfall module using dilated neighborhood attention and reports +1.2 AP over Swin-B on COCO val.","lead":"This paper introduces WTPose, a pose-estimation network that combines a modified Swin transformer with a waterfall module that fuses multi-scale features using dilated neighborhood attention. On COCO validation it reports 77.1 average precision, about 1.2 points above the Swin-B baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 omits the Swin-B + Stem condition without WTM, so the +1.2 AP cannot be attributed to the waterfall transformer; the 0.2 AP edge over ViTPose-B is within typical seed variance. This missing ablation is the load-bearing gap.","rationale":"The reader's weakest assumption focused on configuration selection and run-to-run variance. I agree with that, but the stronger, more internal concern is the missing Stem-only ablation. The paper's attribution of the WTPose gain to the waterfall transformer module (Section 4.1) relies on Table 2, but that table confounds the Stem and WTM contributions. Without a Swin-B + Stem no-WTM condition, the +0.6 AP from row 3 to row 4 could be entirely due to the Stem, and the WTM's independent contribution might be zero or even negative. The figure's architecture and the text suggest the Stem is a meaningful modification; the reader already inferred it contributes about 0.6 AP, but that inference assumes additivity. My concern is that the claim 'waterfall transformer module improves the feature maps' is an interpretation not directly tested. This does not refute the architecture's empirical result as a whole — the full WTPose may indeed outperform the listed baselines — but it does mean the central novelty claim (the waterfall transformer) is not the likely source of the reported gain. Combined with the small absolute margin over ViTPose-B and absent error bars, the paper's conditional acceptance is the right call; however, the missing ablation should be an explicit condition, and the test I propose would resolve the attribution. I do not see an internal inconsistency or theoretical flaw that would justify rejection; the concern is about evidence strength.","tokens_in":6779,"tokens_out":6569,"duration_ms":54036,"concrete_test":"Train and evaluate the four-way ablation under identical mmpose settings: (1) Swin-B baseline, (2) Swin-B + Stem (no WTM), (3) Swin-B + WTM alternating (no Stem), (4) full WTPose. Use three random seeds and report mean±std AP on COCO val. If condition (2) reaches ~77.0 AP, the waterfall transformer contributes little beyond the Stem; if the seed-averaged gap between full WTPose and ViTPose-B falls below ~0.3 AP, the 'outperforms' claim lacks statistical support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the proposed WTPose outperforms other transformer architectures for multi-person pose estimation (Section 5). The evidence rests on Table 1 (AP 77.1 vs 76.9 for ViTPose-B and 75.9/76.3 for Swin-B/Swin-L) and on the ablation in Table 2. Table 2 reports four conditions: Swin-B (75.9), Swin-B + WTM with dilation (2,4,4,8) (76.1), Swin-B + WTM with alternating (2,1,4,1,4,1,8,1) (76.5), and Swin-B + Stem/ResNet-bottleneck + WTM (77.1). Critically, the table never includes Swin-B + Stem without WTM. Consequently, the +0.6 AP jump from row 3 to row 4 cannot be isolated to the Stem; it may be an interaction effect, and the Stem alone may account for most of the improvement. Since the stem is a convolutional front-end (two 3x3 convs plus the first ResNet-101 block), not part of the waterfall transformer, the paper's attribution of the improvement to the waterfall module is not established. Furthermore, the 0.2 AP gap over ViTPose-B (77.1 vs 76.9) is smaller than typical run-to-run variance for COCO pose estimation (often 0.3-0.5 AP), and no standard deviations, multi-seed results, or test-dev numbers are reported. Thus the headline claim is either unsupported or only weakly supported without an additional ablation isolating the WTM from the Stem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes WTPose, a top-down multi-person pose estimation framework built on a Swin transformer backbone with a modified stem (two 3x3 convolutions followed by the first ResNet-101 block) and a Waterfall Transformer Module (WTM). The WTM upsamples and concatenates feature maps from all four Swin stages, then applies a cascade of dilated and non-dilated neighborhood attention transformer blocks with dilation rates (2,1), (4,1), (4,1), (8,1). Experiments on COCO train/val 2017 report AP 77.1 at 89.3M parameters, compared with 75.9 for Swin-B, 76.3 for Swin-L, and 76.9 for ViTPose-B. Table 2 ablates the dilation schedule and the stem/ResNet bottleneck, showing that the alternating schedule improves over uniform dilation and that adding the stem/bottleneck further improves AP. The paper concludes that WTPose outperforms other transformer architectures for multi-person pose estimation.","tokens_in":7133,"tokens_out":5497,"duration_ms":48033,"significance":"If the reported result is reproducible, the contribution is a modest but useful empirical improvement: +1.2 AP over Swin-B and +0.8 AP over Swin-L at 54% fewer parameters than Swin-L, achieved with a relatively simple architectural module. The WTM design is clearly inspired by D-WASP and DiNAT, and the paper is transparent about that lineage. The main limitations are evidential rather than conceptual: the results are from a single validation split with no variance estimates, the ablation table does not isolate the stem effect, and the headline claim is broader than the set of compared baselines. These issues are addressable with additional experiments and careful rewording.","major_comments":[{"comment":"The ablation table does not include a Swin-B + BN condition without the WTM (i.e., the Stem and ResNet bottleneck added to Swin-B only). As a result, the +0.6 AP gain from row 3 (76.5) to row 4 (77.1) cannot be definitively attributed to the stem; it may reflect an interaction between the stem and the WTM. Please add this missing condition so that the stem effect and the WTM effect can be separated. Note that comparing row 1 (75.9) and row 3 (76.5) already supports a WTM-only gain of 0.6 AP, so this gap is about precise attribution rather than the existence of any improvement.","section":"Section 4.1, Table 2"},{"comment":"The claim that WTPose \"outperforms other transformer architectures for multi-person pose estimation\" is broader than the evidence in Table 1, which compares only Swin-B, Swin-L, ViTPose-B, and HRNet (a CNN). Important transformer-based pose estimators such as HRFormer, TokenPose, and TransPose are not evaluated. Please either add these baselines or restrict the claim to the specific transformer baselines compared in the paper.","section":"Abstract and Section 5"},{"comment":"Only one COCO validation split is reported, with no repeated seeds, standard deviations, or test-dev numbers. The 0.2 AP advantage over ViTPose-B (77.1 vs 76.9) is within typical run-to-run variability for COCO pose estimation, so the ranking between WTPose and ViTPose-B is not established without additional runs or a test-dev evaluation. At minimum, report the mean and standard deviation over at least three training runs for the main configurations, and ideally provide a test-dev submission.","section":"Section 4.1, Table 1"}],"minor_comments":[{"comment":"In the sentence following the equations, \"N-MSHA\" appears to be a typo for \"N-MHSA\"; please correct it.","section":"Section 3"},{"comment":"The phrases \"by 1.2%\" and \"by 0.9%\" should be \"by 1.2 AP\" and \"by 0.9 AR\" (or \"percentage points\"), since AP and AR are already reported as percentages.","section":"Section 4.1"},{"comment":"The text says \"summation denotes concatenation,\" but Eq. (1) uses the plus symbol and the text elsewhere says \"summation operator\" while Figure 3 labels the operation \"Concatenate.\" Please use a consistent notation (e.g., \\oplus) and define the concatenation axis explicitly.","section":"Section 3, Eq. (1)"},{"comment":"The Swin-B and Swin-L rows cite reference [7], which is the MMPose codebase; the Swin architecture should be cited as [14] instead.","section":"Table 1"},{"comment":"The sentence \"The contribution of this paper are the following\" has a subject-verb agreement error; it should be \"The contributions of this paper are the following.\"","section":"Section 1"},{"comment":"The decoder used to produce the K heatmaps is not described beyond the figure. Please specify its architecture (e.g., the default MMPose decoder used for Swin-B) so that the method is fully reproducible.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a modest incremental architecture contribution. The core idea is plausible and the reported numbers, if reproducible, would be a useful data point for transformer-based pose estimation. The main risk is the strength of the empirical evidence: the missing stem-only ablation, the lack of multi-seed variance, and the narrow baseline set. I do not see a circularity problem; the design choices are ordinary hyperparameter fitting on the validation set. If the authors supply the missing ablation and additional variance/test-dev evidence, I would be willing to support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines two existing ideas the authors know well, the waterfall multiscale aggregation from their WASP/D-WASP work and dilated neighborhood attention from DiNAT, and applies the combination to Swin for top-down pose estimation. That specific combination is new, and the results are positive: WTPose gets 77.1 AP on COCO val, which is 1.2 AP over Swin-B and 0.8 AP over Swin-L, with only 89.3M parameters. The ablation in Table 2 shows the waterfall transformer module itself contributes about 0.6 AP (75.9 to 76.5), and the stem/bottleneck adds another 0.6. That is a credible engineering result and the paper is honest that the improvements come from the module plus the modified front-end.\n\nWhere the paper gets soft is in what the headline claims. The abstract says the architecture \"outperforms other transformer architectures for multi-person pose estimation,\" but the evidence is a single validation split with no test-dev numbers, no error bars, and no repeated-seed runs. The 0.2 AP edge over ViTPose-B (77.1 vs 76.9) is right in the range of typical run-to-run variance for COCO pose estimation, so that particular comparison is not established. More concretely, the ablation never includes Swin-B plus the stem without the waterfall transformer. Without that row, you cannot tell whether the last 0.6 AP comes from the stem alone or from an interaction with the WTM. That is a genuine gap, but it is fixable, not fatal. The WTM's own contribution is already visible in rows 2 and 3 of Table 2, so the central architectural idea is not unsupported; what is unsupported is the full attribution of the final number.\n\nNo code is released, and the hyperparameters (dilation schedule, stem design) were chosen on the val split. That is ordinary practice, but combined with the missing ablation and narrow baseline set, it means the paper overreaches in its conclusion. I would not call it flawed in a way that justifies rejection. The work is incremental, clearly written, and reproducible in principle if the authors release code and the missing ablation.\n\nWho is this for? People working on transformer backbones for dense prediction, especially pose estimation, who want a cheap way to squeeze a point of AP out of Swin. A serious referee should see it, but the authors should be asked for the missing ablation, multi-seed or test-dev numbers, and ideally code before acceptance.\n\nMy recommendation: send it to peer review, but mark it as conditional on those additions.","headline":"WTPose is a modest but real improvement over Swin for pose estimation; the missing stem-only ablation and thin evaluation make the headline claim too broad.","tokens_in":7661,"tokens_out":1202,"would_cite":true,"duration_ms":12426,"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":"WTPose, a single-pass transformer for multi-person pose estimation, reaches 77.1 AP on COCO validation by feeding multi-scale backbone features through a waterfall of alternating dilated and local window attention blocks, outperforming…","keywords":["human pose estimation","multi-person pose estimation","vision transformer","waterfall transformer module","dilated neighborhood attention","multi-scale features","COCO keypoint detection","Swin transformer"],"falsifier":"Train WTPose and its Swin-B baseline several times with different random seeds under identical training settings and compare AP on a held-out test set; if the average gap drops below roughly 0.5 AP or run-to-run intervals overlap, the consistent-improvement claim loses support. A second check is to replace the alternating dilation schedule with non-dilated attention only: if AP does not fall, the dilation pattern is not the active ingredient.","tokens_in":6566,"feed_emoji":"🦴","tokens_out":6071,"duration_ms":46084,"temperature":0.7,"pith_summary":"WTPose is a single-pass, end-to-end trainable model for top-down multi-person 2D pose estimation. It combines a Swin transformer backbone, with its patch-partition layer replaced by two 3×3 convolutions plus a ResNet-101 residual block, with a transformer-based waterfall module that fuses feature maps from all four backbone stages. The waterfall module runs those features through a cascade of alternating dilated and non-dilated window self-attention blocks, with dilation rates (2,1), (4,1), (4,1), (8,1), which enlarges the receptive field while preserving local detail. On COCO validation the model reaches 77.1 AP, beating the Swin-B baseline by 1.2 AP and the much larger Swin-L by 0.8 AP. If this holds up, it suggests that multi-scale feature fusion plus alternating wide and narrow attention is a cheap way to make transformer backbones better at keypoint localization.","feed_headline":"Waterfall transformer tops Swin on COCO pose at 77.1 AP","feed_subtitle":"A cascade of wide and local attention blocks adds multi-scale context while staying 54% smaller than Swin-L.","key_machinery":"The Waterfall Transformer Module (WTM). It upsamples Swin stage-2, -3, and -4 features to the stage-1 resolution (1/4 scale), concatenates them with low-level ResNet bottleneck features, and reduces channels to 128 with 1×1 convolutions. The resulting tensor flows through four Waterfall Transformer Blocks (WTBs), each pairing a dilated multi-head neighborhood self-attention (D-MHSA) with an MLP and then a non-dilated multi-head neighborhood self-attention (N-MHSA) with an MLP; dilation rates are (2,1), (4,1), (4,1), (8,1) at a window size of 7 with 8 heads. Each WTB branch is concatenated with a depth-wise pooling stream of the input, then low-level features are added and refined with 1×1 and 3×3 convolutions. The mechanism's job is to expand the field of view from 7×7 up to 49×49 while retaining fine local attention, so the decoder gets both global context and local precision.","core_discovery":"The central claim is that waterfall-style multi-scale fusion, implemented with dilated neighborhood attention, gives a consistent performance gain for transformer-based pose estimation. Concretely, WTPose with a modified Swin-B reports AP 77.1, AP75 84.1, AR 82.0 at 89.3M parameters, versus Swin-B at 75.9 AP and Swin-L at 76.3 AP. The ablation in Table 2 separates the gain: adding the waterfall transformer module to Swin-B raises AP from 75.9 to 76.5, and adding the convolutional stem plus ResNet bottleneck raises it further to 77.1. The paper concludes that combining the two components improves feature representation and keypoint accuracy without post-processing.","pith_inferences":["The same waterfall module should transfer to other hierarchical backbones and dense prediction tasks, since it only consumes multi-scale feature maps and attaches at the decoder side; a direct test would be plugging it into an HRNet-style or ConvNeXt-style backbone for semantic segmentation.","Part of the gain may come from the convolutional stem replacing patch partition, a change known to help hierarchical transformers generally; the paper's ablation does not fully separate the stem effect from the waterfall effect beyond the joint table.","A natural extension is to feed the 1/8- or 1/16-resolution features into additional WTM branches, since the current design already uses all four stages but branch count and dilation schedule could be tuned per task rather than fixed.","The approach could be tested on bottom-up pose estimation by running WTM on full-image features without person detection boxes, following the same waterfall idea used in prior atrous-convolution pose models."],"forward_implications":["If the result is right, a Swin-B-sized model (89.3M parameters) can beat Swin-L (197M parameters) by 0.8 AP, making the waterfall module a parameter-efficient accuracy boost.","The alternating dilation schedule (2,1), (4,1), (4,1), (8,1) outperforms dilated-only rates (2,4,4,8), so interleaving wide and narrow attention is a better recipe inside a cascade.","The stem plus ResNet bottleneck front-end contributes roughly 0.6 AP on top of the waterfall module, meaning the input stem matters nearly as much as the multi-scale fusion.","WTM's multi-scale fusion also improves hard cases: AP on medium instances rises to 73.4 and on large instances to 83.9, both above the corresponding Swin baselines."],"supporting_citations":[{"why":"Swin transformer; supplies the backbone architecture, pretrained weights, and the Swin-B/Swin-L baselines that WTPose is compared against.","marker":"[14]"},{"why":"DiNAT dilated neighborhood attention; the building block for the dilated and non-dilated attention used in the waterfall transformer blocks.","marker":"[9]"},{"why":"BAPose/D-WASP; the atrous waterfall representation that the WTM design is explicitly inspired by.","marker":"[3]"},{"why":"Provides the training pipeline, default settings, and evaluation harness used for all experiments, including the Swin baselines.","marker":"[7]"},{"why":"COCO dataset and OKS metric; the benchmark on which all AP and AR numbers are measured.","marker":"[13]"},{"why":"ViTPose; a transformer pose baseline reported in Table 1 as a comparison point.","marker":"[24]"},{"why":"HRNet; a high-resolution CNN pose baseline included in the comparison table.","marker":"[18]"}],"fun_headline_variants":["Waterfall attention lifts COCO pose to 77.1 AP","WTPose beats Swin at 77.1 AP, 54% smaller","Multi-scale waterfall transformer tops Swin on pose","Dilated neighborhood attention adds context, hits 77.1 AP","Single-pass pose net: 77.1 AP with waterfall fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 1.2-point average-precision gain over Swin-B rests on the specific alternating dilation schedule (2,1), (4,1), (4,1), (8,1) and the stem-plus-bottleneck front-end, both chosen after experiments on the COCO validation set; if that configuration is over-tuned to this validation split, or if the gain is within run-to-run variation, the improvement attributed to the waterfall transformer is not established.","fun_headline_variants_meta":{"raw":{"variants":["Waterfall attention lifts COCO pose to 77.1 AP","WTPose beats Swin at 77.1 AP, 54% smaller","Multi-scale waterfall transformer tops Swin on pose","Dilated neighborhood attention adds context, hits 77.1 AP","Single-pass pose net: 77.1 AP with waterfall fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000774,"raw_usage":{"total_tokens":3349,"prompt_tokens":791,"completion_tokens":2558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":2467}},"tokens_in":407,"tokens_out":2558,"duration_ms":17260,"temperature":1.0,"reasoning_tokens":2467,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:42:29.902146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train WTPose and its Swin-B baseline several times with different random seeds under identical training settings and compare AP on a held-out test set; if the average gap drops below roughly 0.5 AP or run-to-run intervals overlap, the consistent-improvement claim loses support. A second check is to replace the alternating dilation schedule with non-dilated attention only: if AP does not fall, the dilation pattern is not the active ingredient.","supporting_citations":[{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows","cited_arxiv_id":null,"evidence_quote":"Swin transformer; supplies the backbone architecture, pretrained weights, and the Swin-B/Swin-L baselines that WTPose is compared against."},{"cited_title":"BAPose: Bottom-up pose estimation with disentangled water- fall representations","cited_arxiv_id":null,"evidence_quote":"BAPose/D-WASP; the atrous waterfall representation that the WTM design is explicitly inspired by."},{"cited_title":"Openmmlab pose estimation toolbox and benchmark","cited_arxiv_id":null,"evidence_quote":"Provides the training pipeline, default settings, and evaluation harness used for all experiments, including the Swin baselines."},{"cited_title":"Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C","cited_arxiv_id":null,"evidence_quote":"COCO dataset and OKS metric; the benchmark on which all AP and AR numbers are measured."},{"cited_title":"Deep high-resolution representation learning for hu- man pose estimation","cited_arxiv_id":null,"evidence_quote":"HRNet; a high-resolution CNN pose baseline included in the comparison table."}],"review_version":1}