{"id":"ed588c99-b080-4327-813d-fcf6f3d9e06d","arxiv_id":"2507.10977","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A wavelet backbone with ray-origin attention encoding improves efficiency and self-comparison accuracy for human-object interaction detection, but remains below the FGAHOI baseline in accuracy despite fewer parameters.","lead":"This paper proposes an image model combining wavelet filters for multi-scale features with learnable ray anchor points that focus attention on important image regions, tested on image classification and human-object interaction detection. It beats its own variants, but trails the FGAHOI baseline by 5.7 mAP on HICO-DET, so its value rests on the new mechanism rather than on state-of-the-art results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III's +3.10 mAP ray gain is not a controlled ablation: Wavelet and Wavelet+3 Rays differ by 2.48M parameters, so the reported gain may not be attributable to the ray mechanism.","rationale":"The reader's verdict is CONDITIONAL, and I agree that conditions should be attached, but I would place the weight differently. The reader's weakest_assumption focuses on the ray mechanism exploiting center-bias in the training data. That is a valid generalization concern and is explicitly acknowledged by the authors in Section V.1. However, the more immediate threat to the central claim is internal: Table III's Wavelet vs Wavelet+3 Rays comparison is not a clean ablation because the parameter counts move in the opposite direction from what appending ray layers would predict. The ImageNet ablation shows ray layers add parameters, while the HOI comparison shows the ray variant has 2.48M fewer parameters. This strongly implies an uncontrolled change in encoder composition, so the quoted +3.10 mAP improvement and the efficiency comparison (34.46M vs 36.94M, 10 vs 9 FPS) are both potentially confounded. This concern is concrete and checkable from the released code or from a simple matched-composition experiment. If the matched-ablation check confirms the gain, the paper's central claim survives and the remaining center-bias caveat is appropriately left as a limitation. If it does not confirm, the central claim is weakened substantially. Since the reader already recommended CONDITIONAL, my verdict remains CONDITIONAL, hence UNCHANGED, but the rationale for the condition should emphasize the matched-composition control as the primary requirement.","tokens_in":11815,"tokens_out":5379,"duration_ms":67677,"concrete_test":"Run HICO-DET with the wavelet backbone and FGAHOI decoder using a fixed total encoder layer count and matched parameter budget for both the Wavelet and Wavelet+3 Rays variants. For example, keep 6 encoder layers total; in the Wavelet row, replace the 3 ray layers with a parameter-matched control such as identity-initialized per-pixel attention or 1x1 convolutional attention, leaving the transformer layers unchanged. If the ray-based variant does not exceed this control by roughly the reported 3.10 mAP, the Table III gain is not attributable to the ray mechanism. Alternatively, inspect the released code to verify the exact layer substitution; if the Wavelet row already uses the same 3 transformer + 3 ray layers with only the ray modules disabled, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim in Section IV.C.1 is that adding ray encoding layers to the wavelet backbone raises HICO-DET mAP from 20.97 to 24.07. Table III, however, shows the two models differ by 2.48M parameters (36.94M vs 34.46M) and by 1 FPS. This is not what one would expect from simply appending ray layers: in the ImageNet ablation (Table II), adding three ray layers increases parameters from 9.58M to 10.38M. The only way to reconcile Table III is that the Wavelet+3 Rays model has a different encoder composition, likely replacing some transformer encoder layers with ray layers, consistent with the implementation detail that the encoder uses '3 ray-based encoding layers and 3 transformer encoder layers.' As a result, the claimed +3.10 mAP gain could be caused by the removed transformer layers, the changed parameter budget, or the ray mechanism itself. Because the paper's central claim rests on this specific self-comparison, the lack of a matched-composition control is the most load-bearing concern. The center-bias issue raised by the reader is real but secondary: a matched ablation would still establish whether the ray module helps on HICO-DET, even if the benefit is a center prior rather than a general attention mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a wavelet attention-like backbone and a ray-based encoder for human-object interaction (HOI) detection. The wavelet backbone is designed to capture middle-order pixel interactions by aggregating features from filters of different sizes, while the ray-based encoder uses learnable ray origins to produce attenuation maps that guide the decoder's attention. The authors evaluate the model on ImageNet-1K for image classification and on HICO-DET for HOI detection. They report that adding three ray layers improves ImageNet top-1 accuracy from 73.36% to 74.54% and improves HICO-DET default mAP from 20.97% to 24.07%, while using fewer parameters and higher FPS than the FGAHOI baseline.","tokens_in":12016,"tokens_out":3755,"duration_ms":45391,"significance":"If the reported gains are attributable to the proposed mechanisms, the paper makes a modest but useful contribution: it demonstrates a parameter-lean alternative architecture for HOI detection and provides one of the first applications of ray-inspired attention to 2D vision. The manuscript is strengthened by the public release of code, by the monotonic improvements when ray layers are added in the ImageNet ablation (Table II), and by a candid discussion of the ray mechanism's data sensitivity in Section V. However, the central HOI ablation is confounded by differences in model composition, and the paper's own analyses show that the ray mechanism encodes a center bias, so the current evidence does not yet support the broader claims of a general multi-scale attention mechanism.","major_comments":[{"comment":"The claimed +3.10 mAP gain from ray-based encoding is not supported as a controlled ablation. Table III reports Wavelet with 36.94M parameters and Wavelet+3 Rays with 34.46M parameters, yet the ImageNet ablation in Table II shows that adding three ray layers increases parameters from 9.58M to 10.38M. The most consistent explanation is that the two HOI models differ in encoder composition, not merely in the presence of ray layers; Section IV.A.2 states that the encoder uses '3 ray-based encoding layers and 3 transformer encoder layers.' To attribute the +3.10 mAP to the ray mechanism, the authors must compare models that differ only in the ray layers, with the backbone, the number of transformer layers, and the parameter budget matched. Without such a control, the gain could be caused by the removed transformer layers or by the changed capacity.","section":"IV.C.1, Table III"},{"comment":"The paper's own evidence indicates that the ray mechanism exploits a center bias rather than providing a general attention prior. Ray origins are initialized on a unit circle because 'most of the objects are centered in the image' (Fig. 5(d)), trained origins converge to the image center (Fig. 5(a-c)), and Section V.1 concedes that the ray layer 'is sensitive to the data' and currently has no directionality. The ImageNet and HICO-DET gains may therefore reflect dataset statistics rather than a reusable spatial-reasoning mechanism. The authors should provide a test on a dataset without a strong center bias, or an explicit spatial-jitter experiment, to substantiate transferability; alternatively, the claims should be narrowed to describe a center-prior module.","section":"III.B.2, IV.B.2, V.1"}],"minor_comments":[{"comment":"Wavelet+1 Ray and Wavelet+2 Rays are both listed with 9.98M parameters, which is surprising because adding a ray layer should change the parameter count; please clarify whether the values are rounded or whether the second ray layer shares parameters with the first.","section":"Table II"},{"comment":"The text calls the ImageNet result 'competitive,' but Table II contains no external baseline such as ResNet, ConvNeXt, or Swin at a comparable parameter and FLOP budget; a direct comparison is needed to support that descriptor.","section":"IV.B.1"},{"comment":"The softmax in Eq. (6) is not fully specified: it is unclear whether normalization is performed over the n ray origins, over the H×W pixel positions, or over both; please state the normalization axis explicitly.","section":"Equation (6)"},{"comment":"The architecture description mentions 'representation learning using 2D Fast Fourier transformation' and 'learning in frequency domain,' but no FFT-based operation appears in the equations or figure; please provide the exact frequency-domain computation or remove the claim.","section":"III.B.2"},{"comment":"The object-center distribution is shown only for ImageNet, yet the ray module is also used on HICO-DET; providing a similar distribution for HICO-DET would help justify the transfer of the center prior.","section":"Fig. 5(d)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of a computer vision journal, but the central claim hinges on a controlled ablation that is currently missing. The authors should be encouraged to run a matched-composition comparison and to either demonstrate transfer of the ray mechanism beyond center-biased datasets or substantially soften the generality claims. The GitHub repository is referenced but was not inspected as part of this review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is the take: the paper introduces a ray-origin attenuation encoder for HOI detection, which is genuinely new in 2D attention, and it is written honestly. But the central quantitative claim—that adding ray layers to the wavelet backbone gives +3.10 mAP on HICO-DET—is not backed by a controlled ablation. The two models differ in encoder composition and parameter count, so the gain cannot be attributed cleanly to the ray mechanism.\n\nWhat is new: the ray mechanism models K and Q using learnable 2D anchor points, PSF spread, and exponential decay, producing a Fourier-domain attention map. That use of ray concepts in HOI attention is absent from the cited prior work, and the design is clearly explained. The wavelet backbone is a variant of existing wavelet CNNs, but the aggregation of low/high-order interactions into middle-order ones is a reasonable framing. Internal ablations on ImageNet are consistent: each added ray layer improves top-1 accuracy, and the paper openly reports the trade-off with FGAHOI (24.07 vs 29.81 mAP) and the efficiency gain (34.46M params, 10 FPS). Section V candidly lists limitations, including the data sensitivity of the ray layer.\n\nThe soft spots, in proportion. The big one is the stress-test concern: Table III's ray gain is confounded. The implementation details say the final model uses \"3 ray-based encoding layers and 3 transformer encoder layers,\" while the Wavelet-only model presumably keeps six transformer layers; that also explains the parameter difference (36.94 vs 34.46M) and the FPS difference. So the +3.10 mAP could come from the removed transformer layers, the changed parameter budget, or the ray mechanism itself. A matched-composition control—same number and width of layers, swapping only the mechanism—is the minimal fix. Second, there are no external baselines on ImageNet and no multiple seeds with variance, so we cannot tell if the improvements are stable. Third, the center-bias limitation is real but secondary; a matched ablation would still show whether the ray mechanism helps on HICO-DET, even if the benefit is a center prior. Minor points: alpha, beta, and PSF initialization are unspecified, and the code repo should have a commit hash for reproducibility.\n\nWho this is for: researchers working on efficient HOI detectors and novel attention priors. The middle-order interaction story is more asserted than measured, but the ray idea is worth examining.\n\nRecommendation: send to peer review. A serious referee should require the matched-composition ablation and variance reporting before the claims can be accepted. If the ray gain survives that, it is a decent contribution; if not, it is a promising idea that currently overstates its evidence.","headline":"A genuinely new ray-based attention mechanism with an honest write-up, but the headline gain on HICO-DET rests on an ablation that is not controlled.","tokens_in":12690,"tokens_out":3024,"would_cite":false,"duration_ms":31226,"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":"This paper claims that a wavelet attention backbone combined with ray-based encoding produces an efficient human-object interaction detector that is leaner and faster than a heavier baseline, at a measured accuracy cost.","keywords":["human-object interaction detection","wavelet backbone","ray-based encoding","multi-scale attention","middle-order interactions","HICO-DET","ImageNet classification","efficient detection"],"falsifier":"Train the same wavelet backbone plus ray encoder on an object-centric dataset whose subjects are not centered, or on HICO-DET after randomly translating objects away from the center; if the mAP gain from rays disappears or reverses, the claimed mechanism is center-bias exploitation rather than general multi-scale attention.","tokens_in":11499,"feed_emoji":"🎯","tokens_out":5279,"duration_ms":58747,"temperature":0.7,"pith_summary":"This paper tries to establish that a backbone built around multi-scale wavelet-like filtering, combined with a ray-based encoding module, can make human-object interaction detection more efficient without collapsing accuracy. The authors claim the wavelet attention stage captures \"middle-order\" interactions between low- and high-order features, which patch-based CNNs tend to miss, and the ray encoder gives the decoder a learnable spatial prior that focuses queries on relevant regions. On HICO-DET the full model reaches 24.07 mAP with 34.46M parameters at 10 FPS, versus 29.81 mAP, 55.42M parameters and 8 FPS for the FGAHOI baseline it builds on. The message is that accuracy-efficiency trade-offs in HOI detection can come from the backbone and attention prior, not only from heavier decoders.","feed_headline":"Ray encoding lifts wavelet HOI detector to 24.07 mAP","feed_subtitle":"Adding learnable ray attention gains 3.10 mAP over the wavelet backbone while cutting parameters versus FGAHOI.","key_machinery":"The two load-bearing components are the wavelet attention backbone and the ray-based encoder. The backbone decomposes feature maps with small and large convolutional filters to produce low- and high-frequency bands, concatenates the four directional sub-bands, then uses a pointwise-convolution modulation computed from these bands to refine features; this is how it aggregates middle-order interactions. The ray encoder treats learnable origin points as attention priors: it computes each pixel's distance to every origin, applies a Gaussian point-spread function and an exponential decay to obtain an attenuation map, passes it through softmax, and uses it to modulate features in the frequency domain via global circular convolution. The learned origins converge toward image centers during training, which the paper interprets as the mechanism learning a center-bias prior.","core_discovery":"The central claim is that aggregating low- and high-frequency convolutional responses into an attention-like modulation recovers middle-order pixel interactions, and that encoding multi-scale attention through learnable ray origins gives a compact way to emphasize task-relevant regions. The paper reports that adding three ray layers to the wavelet backbone raises ImageNet top-1 accuracy from 73.36% to 74.54% and raises HICO-DET default Full mAP from 20.97% to 24.07%, while using fewer parameters and running faster than the FGAHOI detector that supplies the decoder.","pith_inferences":["The ray module's convergence toward image centers suggests it is essentially a learned spatial prior; a natural extension the paper does not test is applying it to datasets where objects are not centered, and if transfer fails there, the mechanism is dataset-specific rather than a general attention prior.","The ray concept could be ported to other dense-prediction tasks such as object detection or segmentation, where a learnable focal region prior might reduce encoder depth while preserving accuracy.","Parameterizing rays with direction or angle, which the paper lists as future work, would let the model represent off-center or elongated attention regions and might close the gap to FGAHOI without adding decoder parameters."],"forward_implications":["If the central claim holds, HOI detectors can be made leaner by replacing heavy encoder stacks with a ray-based attention prior plus a wavelet backbone, without giving up the DETR-style decoder.","Ray encoding transfers from image classification to HOI detection: the same module that improves ImageNet top-1 accuracy also adds 3.10 mAP on HICO-DET default Full.","The wavelet-only variant already reaches 20.97 mAP on HICO-DET, indicating the backbone itself contributes a large share of the final accuracy, with rays contributing the remaining 3.10 points.","Fewer parameters and higher FPS mean the proposed architecture is better suited than FGAHOI to real-time or resource-constrained HOI applications, at a measured 5.74 mAP accuracy cost."],"supporting_citations":[{"why":"Supplies the decoder and training settings, and serves as the parameter-heavy baseline the paper compares against.","marker":"[23]"},{"why":"Provides the convolutional modulation via element-wise multiplication used in the feature refinement stage.","marker":"[39]"},{"why":"Defines the attention formulation that the ray-based encoder adapts into a spatial attenuation prior.","marker":"[40]"},{"why":"Establishes the equivalence between element-wise multiplication and depthwise global circular convolution, which the ray encoder uses for frequency-domain modulation.","marker":"[41]"},{"why":"The Swin Transformer backbone of FGAHOI, used as the comparison point for attention quality and HOI accuracy.","marker":"[14]"},{"why":"The DETR architecture that underlies the one-stage end-to-end HOI detection paradigm the paper builds on.","marker":"[1]"}],"fun_headline_variants":["Wavelet+ray HOI detector hits 24.07 mAP with fewer params","Ray attention boosts wavelet HOI to 24.07 mAP, cuts compute","Multi-scale ray attention lifts HOI detection to 24.07 mAP","Efficient ray-based HOI detector improves ImageNet and HICO-DET"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ray module's benefit comes from a reusable attention prior, not from exploiting the fact that objects tend to sit in the center of the training images.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet+ray HOI detector hits 24.07 mAP with fewer params","Ray attention boosts wavelet HOI to 24.07 mAP, cuts compute","Multi-scale ray attention lifts HOI detection to 24.07 mAP","Efficient ray-based HOI detector improves ImageNet and HICO-DET"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1428,"prompt_tokens":855,"completion_tokens":573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":487}},"tokens_in":471,"tokens_out":573,"duration_ms":6420,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:21:21.482949+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same wavelet backbone plus ray encoder on an object-centric dataset whose subjects are not centered, or on HICO-DET after randomly translating objects away from the center; if the mAP gain from rays disappears or reverses, the claimed mechanism is center-bias exploitation rather than general multi-scale attention.","supporting_citations":[],"review_version":1}