{"id":"e258c446-daf4-4e60-92a7-fcadb7e1e231","arxiv_id":"2502.05147","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LP-DETR proposes a progressive relation-aware self-attention module that models local, medium, and global spatial relations between DETR queries, reaching 52.3 AP with ResNet-50 after 12 epochs on COCO, a small gain over Relation-DETR.","lead":"LP-DETR adds a learnable, layer-wise multi-scale spatial relation module to DETR-style object detectors, reporting small AP gains over Relation-DETR on COCO and two defect-detection datasets. The paper also claims that the model naturally shifts from local to global spatial relations across decoder layers, but this pattern is partially built into the medium-relation scaling that grows with layer index.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 0.3–0.6 AP gains over Relation-DETR are reported without error bars or multiple seeds; they may lie within run-to-run variance, so the central improvement claim is not yet supported.","rationale":"I examined the central claim: LP-DETR improves DETR detection through layer-wise progressive relations. The empirical support rests on 0.3–0.6 AP gaps over Relation-DETR across three settings. These gaps are close to the typical seed-to-seed variability in transformer-based detectors, and the paper reports no variance. This makes the primary quantitative claim the least secure. The reader's identified assumption—that the medium-relation schedule in Eq. 7 biases the observed local-to-global progression—is also valid and directly affects the 'learned pattern' contribution and the attribution of the +0.3 AP from progressive refinement. I therefore partially agree with the reader. My concrete test combines both: multiple-seed evaluation to test significance, and a constant-scale ablation to test whether the progression is learned or imposed. Since the paper provides no code or public artifacts, these checks are essential before the claims can be accepted. The conditional verdict remains appropriate.","tokens_in":9373,"tokens_out":6302,"duration_ms":59492,"concrete_test":"Run the COCO 12-epoch ResNet-50 experiment for LP-DETR and Relation-DETR under identical hyperparameters with 5 random seeds each, and report mean±std AP (and ideally 95% CI). If the intervals overlap by more than one standard deviation, the improvement is not statistically supported. In parallel, ablate Eq. 7's (1+2l/L) schedule by setting the medium scale factor to a constant (e.g., 1.0) or a learned per-layer scalar; if AP changes by less than the seed variance while the Fig. 3a pattern flattens, the 'progressive' claim is an artifact of the fixed schedule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline numbers—52.3 vs 51.7 AP (Table 1), 58.1 vs 57.8 (Table 2), and ablation 52.3 vs 51.1 (Table 5)—are all single-run results. No standard deviations, seed counts, or significance tests are reported. In DETR-style training at 12 epochs, run-to-run AP variance is typically 0.2–0.5 AP, so the reported improvements of 0.3–0.6 AP are comparable to noise. The convergence curves in Fig. 3b also have no error bands. If the differences are not reproducible across seeds, the central claim that LP-DETR improves detection accuracy fails. Additionally, the 'progressive' pattern in Fig. 3a is partly forced: Eq. 7 defines the medium relation R_m with a layer-increasing factor (1+2l/L), so deeper layers receive a different input scale by construction; the local-to-global trend is therefore not purely learned. The paper provides no ablation that removes or alters this schedule. However, the accuracy claim is the foundation, and without statistical support it is the more load-bearing issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"LP-DETR proposes a modification to DETR-style decoder self-attention in which object queries are augmented with multi-scale geometric relation features. The method defines local, medium, and global relation metrics based on log-distances, aspect ratios, and IoU, combines them with per-layer learnable weights Λ, and introduces a layer-dependent scaling factor for the medium relation. Experiments on COCO 2017, CSD, and MSSD report improvements over Relation-DETR (e.g., 52.3 vs. 51.7 AP at 12 epochs with ResNet-50 and 58.1 vs. 57.8 AP with Swin-L), accompanied by ablations on the number of relation heads, the IoU variant, and the progressive refinement module, as well as a qualitative analysis claiming that learned relation weights progress from local to global across decoder layers.","tokens_in":9657,"tokens_out":7013,"duration_ms":63867,"significance":"The core idea of explicit multi-scale geometric relation modeling in DETR self-attention is a reasonable and potentially useful extension of Relation-DETR, and the paper gives a relatively clear module description with several ablations. The paper's strengths include isolating the relation module in Table 5, varying the number of relation heads in Fig. 2a, testing different IoU variants in Fig. 2b, and evaluating on three datasets. However, the reported gains are small (0.3–0.6 AP) and are presented without error bars, seed counts, or significance tests; moreover, the claimed local-to-global progression is partly imposed by the hand-chosen layer-dependent medium-relation scaling in Eq. (7). If the accuracy gains survive multi-seed evaluation and the progression is shown to be learned rather than prescribed, this would be a modest but solid incremental contribution. In its current form, the evidence does not yet support the headline claims.","major_comments":[{"comment":"Every claimed improvement over Relation-DETR (Table 1: +0.6 AP at 12 epochs and +0.4 AP at 24 epochs; Table 2: +0.3 AP; Tables 3–4: +0.4 and +0.6 AP; Table 5: +0.9 AP for the relation module and +0.3 AP for progressive refinement) is based on a single run with no standard deviations, seed counts, or significance tests. DETR-style detectors trained for 12 epochs typically exhibit run-to-run AP variation of 0.2–0.5 AP, so the reported gains are comparable to plausible noise. The central accuracy claim is therefore not yet supported; please provide mean and standard deviation over at least three seeds for the main comparisons and the key ablations, or otherwise demonstrate that the differences are reproducible.","section":"§4.2, Tables 1–5"},{"comment":"The claimed local-to-global progression is not a purely learned phenomenon. The medium-relation metric is defined with the layer-dependent scaling factor (1+2l/L), so the input to the learned weights Λ changes systematically with decoder depth before any training. The paper does not ablate this hand-chosen schedule, for example by fixing the factor to 1, removing it, or using an inverted schedule. Without such an ablation, the claim that the model 'naturally learns to prioritize local spatial relations in early decoder layers' is at least partly an artifact of the definition, and the accuracy contribution of the progressive component in Table 5 (+0.3 AP) cannot be attributed to learned progression.","section":"§3.2, Eq. (7) and §4.4"}],"minor_comments":[{"comment":"The Swin-L result is reported as 58.0 AP in the Abstract and Introduction but as 58.1 AP in Table 2; please reconcile the headline number.","section":"Abstract, Section 1, and Table 2"},{"comment":"The subscript and superscript notation in the equations is difficult to parse (for example, Eq. (1) mixes q_i^c, q_i^p, and q_i^r with an unclear summation index), and the dimensions of the projection matrices W_q, W_k, W_v, and W_g are not stated. Please rewrite the notation cleanly and define all tensor shapes.","section":"Equations (1)–(7)"},{"comment":"Relation-DETR is cited as [11] in the section text but appears as [10] in the reference list, where [11] is Salience-DETR; please correct the citation labels.","section":"Section 2.1 and reference list"},{"comment":"For ResNet-50 at 12 epochs, Table 1 shows AP_S decreasing from 36.1 (Relation-DETR) to 35.8 (LP-DETR), so the conclusion's statement of 'marginal improvements on small objects' is not supported by the reported numbers; please report and discuss this decrease.","section":"§4.2 and Conclusion"},{"comment":"The analysis of local-to-global relation weights should state how the plotted values are normalized and what the y-axis represents; without this, statements such as 'weights exceeding 0.9' and 'decrease to 0.24' cannot be verified.","section":"§4.4 and Fig. 3a"},{"comment":"The statement that LP-DETR achieves 'state-of-the-art performance of 58.1 AP' is based on only four Swin-L baselines; please either expand the comparison to a broader set of recent Swin-L DETR detectors or qualify the claim as improvement over the listed baselines.","section":"§4.2, Table 2"},{"comment":"CSD and MSSD are described as having 373 and 962 training images, respectively; please clarify whether these are the complete dataset sizes and specify the evaluation splits, since the small scale of these datasets makes single-run comparisons especially noisy.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"This is an incremental DETR variant with a plausible mechanism, but the present evidence is thin: all accuracy comparisons are single-run, and the central interpretative claim of a learned local-to-global progression is entangled with the hand-chosen layer-dependent scaling in Eq. (7). I would advise major revision rather than rejection, because the missing experiments—multi-seed evaluation and an ablation of the scaling schedule—are well-defined and within the scope of a revision. I also recommend that the editors ask the authors to resolve the 58.0/58.1 inconsistency before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent incremental paper. The new bit is a layer-wise learned weighting over three hand-defined spatial relation scales (local, medium, global) injected into DETR decoder self-attention. It is a natural extension of Relation-DETR and Relation Networks, and the authors are transparent about that lineage. What the paper does well: the module is clearly specified, the main comparison is against the closest baseline, ablations are consistent, and the two defect datasets give some breadth. The learned local-to-global story is a nice readout from Fig. 3a.\n\nThe soft spot is statistical. Every reported AP difference—52.3 vs 51.7, 58.1 vs 57.8, the 0.9/0.3 ablation steps—comes from a single run. There are no seeds, no error bars, no significance test. At 12-epoch DETR training, run-to-run AP spread of 0.2–0.5 is normal, so the 0.3–0.6 gains are inside the noise band. That does not mean the method is fake; it means the central claim is not yet supported. The progressive pattern is also partly built in: the medium relation scale multiplies by (1+2l/L), so deeper layers get a different relation metric by construction. No ablation removes that schedule, so the “naturally learns local-to-global” claim is weaker than presented.\n\nMinor issues: abstract says 58.0 AP for Swin-L while Table 2 says 58.1; “state-of-the-art” is used with a short baseline list; CIoU is chosen on a marginal difference. None of these are load-bearing. The citation pattern is fine; Relation-DETR and Relation Network are properly credited.\n\nFor a reader working on DETR relation modeling, this is a useful incremental data point, worth a reading group slot. For peer review, send it out: the method is clearly specified and the experiments are ablatable, but ask for multiple seeds/error bars and an ablation with a fixed or reversed medium-relation schedule. With those, a conditional accept is defensible; without them, the headline gain is unverified.","headline":"A clean, clearly specified incremental extension of Relation-DETR, but the headline accuracy gains are single-run numbers and could easily sit inside run-to-run noise.","tokens_in":10191,"tokens_out":2455,"would_cite":false,"duration_ms":24779,"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":"LP-DETR claims that adding learnable multi-scale spatial relations to decoder self-attention, with the relation mix shifting from local to global across layers, improves DETR-based object detection.","keywords":["object detection","DETR","transformer decoder","self-attention","spatial relations","multi-scale relations","layer-wise progressive","COCO"],"falsifier":"A reader could settle the claim by training LP-DETR with the medium-relation scaling factor held constant across all six decoder layers (and, as a second arm, reversed so early layers see broad relations and later layers see narrow ones) while keeping the relation-aware attention module unchanged. If AP stays at 52.3 or the layer-wise weight curve still runs local-to-global, the progressive schedule is not the cause of the gain; if AP drops toward the 51.1 baseline of relation-aware attention without progressive refinement, the schedule is load-bearing.","tokens_in":1897,"feed_emoji":"🎯","tokens_out":5292,"duration_ms":98519,"temperature":0.7,"pith_summary":"This paper argues that DETR-style object detectors leave a useful signal unused: the spatial relations between object queries inside decoder self-attention. The authors propose LP-DETR, which augments each query with a relation-aware attention term built from geometric relation metrics — log relative distances, scale ratios, and IoU between reference boxes — at three scales: local, medium, and global. A learnable per-layer weight balances the three scales, and the medium scale is deliberately broadened as the decoder gets deeper. On COCO 2017 the method reaches 52.3 AP with ResNet-50 at 12 epochs and 52.5 AP at 24 epochs, exceeding its direct predecessor Relation-DETR by 0.4–0.6 AP, and 58.1 AP with a Swin-L backbone. The authors further report that the learned weights shift from local-dominated early layers to medium/global-dominated deep layers, which they interpret as evidence that detection benefits from hierarchical spatial reasoning.","feed_headline":"Layer-wise spatial relations push DETR to 52.3 AP on COCO","feed_subtitle":"Relation-aware self-attention with layer-wise local-to-global scaling beats Relation-DETR by up to 0.6 AP.","key_machinery":"The load-bearing object is the layer-wise progressive relation-aware self-attention module inserted into the DETR decoder's self-attention. For each object query it computes a relation query as a weighted sum of content queries, with weights that combine ordinary attention scores with a geometric weight; the geometric weight is produced by an MLP over sinusoidal embeddings of local, medium, and global relation metrics derived from reference-box coordinates (Eqs. 1–7). The progressive part is the medium relation metric, whose log-scale input is divided by $(1 + 2l/L)$ at decoder layer $l$, making relations broader in later layers, together with a learnable per-layer vector $\\Lambda = [\\lambda_{\\text{local}}, \\lambda_{\\text{medium}}, \\lambda_{\\text{global}}]$ that rebalances the three scales. This object carries the argument because it is the only new component added to the baseline detector: the gains are attributed to it, and the observed local-to-global weight trajectory is read off from the learned $\\Lambda$ values.","core_discovery":"On the paper's own terms, the central discovery is that explicitly modeling multi-scale spatial relationships between object queries in the decoder's self-attention improves both convergence speed and final detection accuracy, and that the model learns a layer-wise progression from local to global relations. The mechanism works by enriching each object query with a relation query: a weighted sum of all content queries, where the weights multiply ordinary attention scores by a geometric weight. The geometric weight is an MLP applied to a sinusoidal embedding of a five-dimensional relation metric (log distance ratios, log width/height ratios, and IoU of the reference boxes), evaluated under three relation scales. The local scale is the raw metric, the global scale uses constant weights, and the medium scale applies a factor of $(1 + 2l/L)$ that grows with decoder layer index $l$, so deeper layers see progressively broader spatial neighborhoods. Results on COCO 2017, CSD, and MSSD show consistent gains over Relation-DETR, and an ablation attributes about 0.9 AP to the relation mechanism and another 0.3 AP to the progressive refinement schedule.","pith_inferences":["In our reading, the reported local-to-global progression is partly imposed by construction: the medium-relation scaling factor $(1 + 2l/L)$ grows with layer index before any training, so the learned $\\Lambda$ weights may simply be amplifying a schedule the architecture already pushes toward. A fair test would hold the schedule fixed or reverse it and check whether the pattern and the +0.3 AP from ","Because the relation module lives entirely in decoder self-attention, it is orthogonal to denoising-based training (e.g., DN-DETR, DINO) and one-to-many matching strategies; combining LP-DETR with those could plausibly stack gains, but the paper does not test this.","The relation metric is defined over reference boxes, so the same mechanism could transfer to other set-prediction tasks that use DETR-style decoders — panoptic segmentation, multi-object tracking, or open-vocabulary detection — wherever box coordinates are available. The paper does not explore these.","If the accuracy gain is driven mainly by the geometric relation prior rather than by the layer-wise progression, a simpler fixed mixture of local, medium, and global relations might achieve most of the benefit; that hypothesis is testable by ablating $\\Lambda$ to constant values."],"forward_implications":["Replacing standard decoder self-attention with relation-aware self-attention raises AP from 51.1 to 52.0 on COCO 2017 with ResNet-50 at 12 epochs; adding the progressive refinement schedule raises it further to 52.3.","Allocating more attention heads to relation-aware computation improves accuracy monotonically: zero relation heads gives 51.1 AP, while eight relation heads give 52.3 AP.","The gain over Relation-DETR is consistent across schedules and backbones: +0.6 AP at 12 epochs, +0.4 AP at 24 epochs with ResNet-50, and +0.3 AP with Swin-L, and the method also leads on the CSD and MSSD datasets.","The learned relation weights show a stable pattern: early decoder layers favor local relations (weights above 0.9), while later layers shift toward medium and global relations, with medium/global weights overtaking local by the final layer.","The model's convergence curve stays above Relation-DETR's throughout training, consistent with the claim that spatial relation modeling speeds up DETR training."],"supporting_citations":[{"why":"Relation-DETR supplies the explicit position-relation prior that LP-DETR extends and is the main baseline it must beat.","marker":"[10]"},{"why":"DETR establishes the end-to-end transformer detection architecture whose convergence problem motivates the work.","marker":"[3]"},{"why":"DAB-DETR provides the decoupled content/position query formulation and reference-box parametrization used in the relation metric.","marker":"[22]"},{"why":"Relation Networks supply the original geometric-weight mechanism for injecting spatial relations into attention.","marker":"[12]"},{"why":"DINO is a strong denoising-anchor baseline and comparison point in the main COCO tables.","marker":"[32]"},{"why":"The standard attention paper supplies the self-attention formulation and sinusoidal encoding used in the relation embedding.","marker":"[28]"},{"why":"Deformable DETR is a multi-scale feature baseline in the comparison and part of the related-work context for convergence speed.","marker":"[37]"}],"fun_headline_variants":["LP-DETR: layer-wise relations push DETR to 52.3 AP","Relation-aware self-attention lifts DETR to 52.3 AP on COCO","Progressive relations make DETR converge faster and sharper","Local-to-global query relations boost DETR's COCO AP to 52.3"],"cache_read_input_tokens":12288,"weakest_assumption_plain":"The load-bearing premise is that the rule that later decoder layers use progressively broader relations — not just the presence of multi-scale geometric relations themselves — is what improves detection; if a fixed or reversed schedule kept the same accuracy and weight pattern, the paper's central explanation would collapse.","fun_headline_variants_meta":{"raw":{"variants":["LP-DETR: layer-wise relations push DETR to 52.3 AP","Relation-aware self-attention lifts DETR to 52.3 AP on COCO","Progressive relations make DETR converge faster and sharper","Local-to-global query relations boost DETR's COCO AP to 52.3"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1754,"prompt_tokens":959,"completion_tokens":795,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":706}},"tokens_in":575,"tokens_out":795,"duration_ms":7810,"temperature":1.0,"reasoning_tokens":706,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T20:03:48.638046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the claim by training LP-DETR with the medium-relation scaling factor held constant across all six decoder layers (and, as a second arm, reversed so early layers see broad relations and later layers see narrow ones) while keeping the relation-aware attention module unchanged. If AP stays at 52.3 or the layer-wise weight curve still runs local-to-global, the progressive schedule is not the cause of the gain; if AP drops toward the 51.1 baseline of relation-aware attention without progressive refinement, the schedule is load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Relation-DETR supplies the explicit position-relation prior that LP-DETR extends and is the main baseline it must beat."},{"cited_title":"=#𝑤!#\"$ #⋅&𝑊%⋅𝑞#( (1) The relation weight 𝑤'-* captures both geometric and attention-based relationships be-tween queries, which is computed as: 𝑤!#","cited_arxiv_id":null,"evidence_quote":"DETR establishes the end-to-end transformer detection architecture whose convergence problem motivates the work."},{"cited_title":"et al.: DETRs with Hybrid Matching","cited_arxiv_id":null,"evidence_quote":"DAB-DETR provides the decoupled content/position query formulation and reference-box parametrization used in the relation metric."},{"cited_title":"et al.: End-to-End Object Detection with Transformers","cited_arxiv_id":null,"evidence_quote":"Relation Networks supply the original geometric-weight mechanism for injecting spatial relations into attention."},{"cited_title":"et al.: Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows","cited_arxiv_id":null,"evidence_quote":"DINO is a strong denoising-anchor baseline and comparison point in the main COCO tables."},{"cited_title":"et al.: Core-Text: Improving Scene Text Detection with Contrastive Relational Rea-soning","cited_arxiv_id":null,"evidence_quote":"The standard attention paper supplies the self-attention formulation and sinusoidal encoding used in the relation embedding."},{"cited_title":"et al.: Deformable DETR: Deformable Transformers for End-to-end Object Detec-tion","cited_arxiv_id":null,"evidence_quote":"Deformable DETR is a multi-scale feature baseline in the comparison and part of the related-work context for convergence speed."}],"review_version":1}