{"id":"483c24e0-a02d-4306-93e0-bbb99a8949a7","arxiv_id":"2507.19807","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A flexible-query transformer detector that separates cross-attention localization from self-attention deduplication reports consistent accuracy gains over DINO across five backbones.","lead":"DS-Det is a transformer-based object detector that adapts how many detection queries it uses for each image instead of always using a fixed number. It separates the decoder into a box-locating stage and a duplicate-removing stage, and reports higher accuracy than DINO on COCO and WiderPerson while running the decoder faster.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FLET threshold S=0.02 is assumed transferable across domains, but only COCO S-sweeps are shown; a calibration shift could invert the accuracy/efficiency claim.","rationale":"The reader's weakest assumption is exactly the load-bearing point. The paper's novelty claim rests on converting fixed queries to flexible ones via a thresholded classifier. The evidence for robustness (Table 14) is single-dataset and does not include calibration. The ablation S=0.01 to 0.05 on COCO shows stability on that distribution, but a detector's score distribution is known to shift with object density, scale, and category priors; WiderPerson has 29.87 annotations/image versus roughly 7 for COCO, so the same threshold can behave very differently. The paper does run WiderPerson experiments but does not report query counts or S sensitivity there, so we cannot tell whether the default S is near-optimal or just acceptable. Since the headline comparison is +1.5 AP, a modest miscalibration could erase it. Other concerns (missing alpha, code commit hash, unquantified ROT) are real but secondary: alpha affects a loss weighting term, code hash affects reproducibility, and ROT is an explanatory story, while the threshold directly implements the central 'flexible' contribution. Therefore the verdict should remain CONDITIONAL: the architecture is plausible and well-ablated, but the flexibility claim is conditional on threshold portability. A single calibration/sensitivity experiment on a second dataset would settle it.","tokens_in":29712,"tokens_out":5194,"duration_ms":66731,"concrete_test":"Take the released DS-Det-ResNet50 checkpoint and run the COCO-trained model on WiderPerson validation with S = 0.01, 0.02, 0.05, 0.10, and 0.20, recording AP, Recall, and mean query count per image. Then compare the optimal S on WiderPerson to the COCO-optimal S. Also compute the expected calibration error (ECE) of the FLET classification head on both COCO val and WiderPerson val. If the WiderPerson AP at S=0.02 is more than 1 AP below the best S on that dataset, or ECE exceeds about 0.1 on either set, the fixed-threshold assumption is not portable and the flexibility claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 'flexible query' mechanism rests on a fixed threshold S=0.02 applied to FLET classification scores (Sec. 3.3, Appendix Table 14). The only sensitivity evidence is a COCO val sweep (Table 14), where AP varies only 50.2-50.6 for S in [0.01,0.05]. This does not establish that the score distribution is calibrated across domains. On WiderPerson or CrowdHuman, the same S is used without reporting AP-vs-S curves; if the encoder classification head's score distribution shifts (e.g., dense pedestrian scenes produce higher background scores), S=0.02 may select too many low-information queries (hurting efficiency) or too few object queries (hurting recall). The paper does not report expected calibration error or any reliability diagram for the FLET head, nor does it ablate S on WiderPerson/CrowdHuman. Since the headline gain over DINO is +1.5 AP at 12 epochs, a domain shift that changes the optimal S by more than roughly 1 AP could reverse the claimed advantage. This is load-bearing because the flexibility and efficiency advantages are exactly what distinguishes DS-Det from fixed-query DETR baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DS-Det, a DETR-family detector that replaces fixed content/positional queries with a flexible set of queries produced by a FLET module (encoder-token classification followed by threshold selection S), partitions the decoder into a Box Locating Part (cross-attention with one-to-many matching) and a Deduplication Part (self-attention with one-to-one matching), and uses a PoCoo loss that reweights classification by box size. Experiments on COCO, WiderPerson, and CrowdHuman across five backbones report consistent AP gains over DINO (e.g., +1.5 AP at 12 epochs on COCO with ResNet50), a 34.8% decoder speedup at 900 queries, and an additional Deformable-DETR variant in the appendix.","tokens_in":29968,"tokens_out":7816,"duration_ms":96980,"significance":"If the empirical results hold, the paper makes a useful contribution: it shows that a decoder with fewer self-attention layers and a single query type can match or beat a fixed-query mixed-attention decoder on equal training budgets, and it provides extensive ablations (component, loss, SA count, threshold, layer split, matching order, cost weights), speed measurements, and cross-backbone/cross-dataset experiments. The FLOPs and query-count analysis, the COCO object-number subsets, and the downstream open-ended detection complexity test are concrete evidence for the flexibility claim. The main risks are the threshold-based query selection's transferability and a train/inference mismatch in query construction, both of which affect the central 'flexible query' claim.","major_comments":[{"comment":"The flexible-query mechanism is controlled by the FLET threshold S=0.02, but the only sensitivity analysis is on COCO val, where AP moves only from 50.2 to 50.6 for S in [0.01,0.05]. The same fixed S is used without additional analysis on WiderPerson (Table 4) and CrowdHuman (Table 20), and no query-count statistics or calibration evidence are reported for these datasets. Because the paper's accuracy/efficiency claim is specifically that FLET selects the right number of queries per image, a score-distribution shift in a new domain could select too many or too few queries and reverse the reported gains; the WiderPerson margins (e.g., +0.49 AP for ResNet50 in Table 4) are small enough that this is a realistic risk. Please provide per-dataset AP-vs-S sweeps (or an alternative adaptive selection criterion) and report the distribution of selected query counts on WiderPerson and CrowdHuman.","section":"Sec. 3.3 and Appendix B.4 (Table 14)"},{"comment":"The paper does not specify the training objective of the FLET classification head or how the sampled queries used during training relate to the thresholded queries used at inference. Table 5 attributes +4.6 AP to FLET over the flexible-c baseline, so this mechanism is load-bearing; however, the reader cannot tell whether the improvement comes from threshold selection, query-sampling regularization, or placeholder-query padding. Please state the classification loss and label assignment for encoder tokens, and ablate query sampling against direct top-k/threshold selection during training.","section":"Sec. 3.3 (Query Sampling and Query Alignment)"},{"comment":"The text states that DS-Det obtains overall higher performance than DINO variants on CrowdHuman, but Table 20 shows DS-Det-Strip-MLP-T at 87.92 AP versus DINO-Strip-MLP-T at 88.38 AP, a decrease of 0.46 AP. Since the paper's general-effectiveness claim rests on consistent improvements, this exception must be acknowledged and explained (or the claim qualified) rather than glossed over.","section":"Appendix B.4 (CrowdHuman, Table 20)"}],"minor_comments":[{"comment":"The second summation in Eq. (2) uses j=i-T1 with the condition i>T1, reusing i from the first summation; this indexing is confusing and should be rewritten with separate layer indices for the BLP and DP blocks.","section":"Sec. 3.2, Eq. (2)"},{"comment":"The figure caption contains the placeholder text '替换图片：', which appears to be an editing artifact and should be removed before publication.","section":"Fig. 2"},{"comment":"The PoCoo loss uses t_i in the BCE term without defining it; the paper should specify how the IA-BCE label from Align-DETR is constructed, since the formula is otherwise incomplete.","section":"Sec. 3.5, Eq. (6)"},{"comment":"The FLOPs for DS-Det-ResNet50 are reported as 275G in Table 3 but as 273G in Appendix C.4 for the same configuration; the numbers should be reconciled.","section":"Table 3 vs. Appendix C.4"},{"comment":"The row label 'flexible-c' in Table 5 is not defined in the table caption; please clarify whether it means a fixed number of queries chosen manually for flexible-conditioned training.","section":"Table 5"},{"comment":"The 'ROT' and 'query ambiguity' constructs are asserted rather than directly measured; the ablation results are consistent with the proposed story, but a direct diagnostic (e.g., query-similarity or gradient-conflict analysis) would strengthen the causal claim that the decoder design addresses these specific phenomena.","section":"Sec. 3.1 and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent empirical study with unusually detailed ablations, and the threshold-transferability concern is addressable with additional experiments. The CrowdHuman reporting inconsistency should be corrected before acceptance; I do not see grounds for rejection, but the central flexibility claim needs the requested per-dataset threshold analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read. The paper does something real: it replaces fixed queries with a threshold-based selection from encoder tokens, and it restructures the decoder so cross-attention localizes (one-to-many) and self-attention deduplicates (one-to-one). The empirical work is unusually broad — five backbones, three datasets, a Deformable-DETR variant, speed tests, and a query-count-versus-object-count analysis. The +1.5 AP over DINO at 12 epochs is credible, and the per-component ablations mostly support the design claims.\n\nThe soft spots are real but not fatal. The biggest is the threshold S=0.02: the paper ablates it on COCO val (Table 14) and shows a robust plateau between 0.01 and 0.05, but it never reports an AP-vs-S curve on WiderPerson or CrowdHuman. The stress-test concern is fair — if the encoder classification head's score distribution shifts in dense pedestrian scenes, the same S might select too many or too few queries. That said, the accuracy gains on those datasets are consistent across four backbones, so even with a suboptimal S the method isn't worse; the efficiency advantage could erode, but the accuracy claim likely survives. A two-line calibration plot or a small S sweep on WiderPerson would close this out.\n\nTwo other things bug me. The \"ROT\" and \"query ambiguity\" constructs are motivating stories, not measured quantities — no direct evidence that the opposing-gradient conflict is what slows down DINO. And the PoCoo loss leaves the exponent alpha unspecified; the ablation table shows the loss helps, but I can't reproduce it from the text. The code is linked but unversioned. These are minor-to-moderate, not load-bearing.\n\nThe paper is for anyone working on DETR-style detectors, flexible/efficient decoding, or small-object performance. It deserves a serious referee; the missing calibration and alpha value are revision items, not grounds for rejection. My verdict: engage with it, ask for the cross-domain S sensitivity and the alpha value, and expect a clean resubmission.","headline":"A solid, carefully ablated DETR variant with a genuinely flexible query mechanism; the central claims hold, though the fixed threshold S deserves one more cross-domain sensitivity test.","tokens_in":30510,"tokens_out":1310,"would_cite":true,"duration_ms":19218,"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":"DS-Det argues that separating the decoder's cross-attention (locating) from self-attention (deduplication), together with image-dependent query selection, lets a transformer detector beat the fixed-query DINO baseline by +1.5% AP on COCO…","keywords":["flexible object detection","single-query paradigm","attention disentangled learning","query ambiguity","DETR decoder","small object detection","efficient transformer decoder"],"falsifier":"Measure the calibration of the FLET classification head on a held-out domain (for example, aerial imagery with hundreds of small objects per image) and compare AP at S=0.02 against an oracle that selects the optimal query count per image; if the fixed threshold consistently picks too few or too many queries, the flexibility claim reverses.","tokens_in":29507,"feed_emoji":"🎯","tokens_out":4807,"duration_ms":50863,"temperature":0.7,"pith_summary":"The paper claims that standard DETR-style detectors hurt themselves by interleaving two opposing attention operations layer by layer: self-attention pushes queries apart while cross-attention pulls them together around the same object, and mixing one-to-one with one-to-many label assignments in shared decoder layers makes each query unsure whether to predict a single object or one of many. To remove this 'query ambiguity', DS-Det lets the encoder choose a flexible number of queries per image and splits decoding into a box-locating stage that uses only cross-attention and a deduplication stage that uses self-attention. If correct, a simpler decoder can be both more accurate and more efficient than existing fixed-query designs, with gains concentrated on small objects and with the query count adapting to the image. The paper demonstrates this on COCO, WiderPerson, and CrowdHuman across five backbones and also shows the recipe transfers to Deformable DETR.","feed_headline":"Split attention roles to boost DETR detection by 1.5 AP","feed_subtitle":"Separating localization from deduplication lets the decoder pick queries per image and run about 35% faster.","key_machinery":"The load-bearing object is the Attention Disentangled Decoder (ADD), built on the new Single-Query paradigm. FLET derives the queries by applying a classification head to all encoder tokens and selecting those above a threshold S, so the number of queries is flexible and tied to image content rather than a fixed hyperparameter. ADD then separates the attention roles: the Box Locating Part contains only cross-attention and uses one-to-many matching (each ground-truth box replicated K times) to locate objects, while the Deduplication Part contains multi-head self-attention blocks and uses one-to-one matching to suppress duplicate predictions, with a stop-gradient preventing the two matching schemes from conflicting. A third component, the PoCoo loss, reweights the classification loss by box size so small objects receive higher learning priority.","core_discovery":"The central discovery is that inefficient training and weak decoders in DETR-like detectors trace back to two entangled sources: the recurrent opposing interactions between cross-attention and self-attention, and the mixing of one-to-one and one-to-many matching within shared-weight decoder layers. DS-Det resolves both by introducing a Flexible Single-Query generation (FLET) module, which runs a classification head over all encoder tokens and uses a threshold to form an image-dependent pool of queries, and by redesigning the decoder as an Attention Disentangled Decoder (ADD). In ADD, the Box Locating Part uses only cross-attention with one-to-many matching to locate objects, while the Deduplication Part uses self-attention with one-to-one matching to suppress duplicates, with a stop-gradient between the two parts so their supervision does not conflict. On COCO, DS-Det with ResNet50 outperforms the DINO baseline by +1.5% AP and +2.3% AP_S under a 12-epoch schedule, improves decoder inference speed by 34.8% at 900 queries, and scales its query count with the number of objects in the image (about 65 queries on sparse images, up to 900 on dense ones).","pith_inferences":["The paper's diagnosis of 'query ambiguity' suggests a general design rule for set-prediction decoders: whenever one-to-many and one-to-one supervision share weights, separate them by stage or by gradient, not just by loss weighting.","Because the threshold S controls the query budget at inference without retraining, DS-Det could be deployed with an adjustable speed-accuracy trade-off, for example raising S on resource-constrained devices or lowering it for dense scenes.","The flexible query count is especially promising for open-ended or language-coupled detection, where feeding 65 rather than 900 queries to a language model would cut downstream computation by roughly an order of magnitude.","A testable extension is applying FLET and ADD to panoptic or video detection, where the number of objects per frame is even more variable and fixed-query decoders waste computation on empty frames."],"forward_implications":["A decoder with only 4 cross-attention-only layers plus 2 deduplication layers can outperform a 6-layer mixed-attention decoder on the same 12-epoch training budget.","Sparse images need far fewer queries: the model uses roughly 65 queries on images with 1-5 objects versus the fixed 900 in DINO, with comparable or better AP.","Decoder inference is faster (9.2 ms vs 14.1 ms at 900 queries), because self-attention layers are concentrated in only two of the six decoder stages.","Small-object AP improves (+2.3% AP_S over DINO at 12 epochs), which the paper attributes to the box-size-aware PoCoo loss.","The recipe is portable: applying FLET, ADD, and PoCoo to Deformable DETR raises its AP by +3.1% at 30 epochs versus the original 50-epoch result."],"supporting_citations":[{"why":"DINO is the primary baseline and comparison model throughout the paper, providing the fixed-query decoder whose attention layers are ablated and whose performance is improved upon.","marker":"[51]"},{"why":"DAC-DETR supplies the prior observation that cross-attention gathers queries while self-attention disperses them, which the paper reformulates as the ROT issue.","marker":"[19]"},{"why":"Align-DETR contributes the IA-BCE loss that PoCoo extends with a box-size prior.","marker":"[2]"},{"why":"Deformable DETR provides the multi-scale deformable cross-attention used in ADD and serves as the second architecture for the generalization experiments.","marker":"[56]"},{"why":"DiffusionDet is the main flexible-object-count comparison, which DS-Det outperforms by a large margin at fewer training epochs.","marker":"[9]"},{"why":"DQ-DETR is the 'classified' baseline that predicts query numbers from fixed bins, contrasted with true continuous flexible selection.","marker":"[20]"},{"why":"DETR establishes the one-to-one matching principle that DS-Det restores by decoupling it from one-to-many matching.","marker":"[7]"},{"why":"Hybrid-DETR represents the one-to-many-matching approach whose added branches and shared weights are cited as sources of query ambiguity.","marker":"[21]"}],"fun_headline_variants":["Split attention roles in DETR decoder: +1.5 AP, 35% faster","Flexible query counts and disentangled attention improve DETR","Single-query DETR: dynamic queries resolve training conflicts","Disentangle self- and cross-attention to fix DETR query ambiguity","DETR: disentangle attention for flexible queries and speed gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the FLET classification head's scores reliably separate object-bearing encoder tokens from background, so a fixed threshold S=0.02 chooses roughly the right number of queries for any image.","fun_headline_variants_meta":{"raw":{"variants":["Split attention roles in DETR decoder: +1.5 AP, 35% faster","Flexible query counts and disentangled attention improve DETR","Single-query DETR: dynamic queries resolve training conflicts","Disentangle self- and cross-attention to fix DETR query ambiguity","DETR: disentangle attention for flexible queries and speed gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000783,"raw_usage":{"total_tokens":3546,"prompt_tokens":1120,"completion_tokens":2426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":736,"completion_tokens_details":{"reasoning_tokens":2328}},"tokens_in":736,"tokens_out":2426,"duration_ms":21681,"temperature":1.0,"reasoning_tokens":2328,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:00:10.821007+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the calibration of the FLET classification head on a held-out domain (for example, aerial imagery with hundreds of small objects per image) and compare AP at S=0.02 against an oracle that selects the optimal query count per image; if the fixed threshold consistently picks too few or too many queries, the flexibility claim reverses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DAC-DETR supplies the prior observation that cross-attention gathers queries while self-attention disperses them, which the paper reformulates as the ROT issue."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Align-DETR contributes the IA-BCE loss that PoCoo extends with a box-size prior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DiffusionDet is the main flexible-object-count comparison, which DS-Det outperforms by a large margin at fewer training epochs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DQ-DETR is the 'classified' baseline that predicts query numbers from fixed bins, contrasted with true continuous flexible selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DETR establishes the one-to-one matching principle that DS-Det restores by decoupling it from one-to-many matching."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Hybrid-DETR represents the one-to-many-matching approach whose added branches and shared weights are cited as sources of query ambiguity."}],"review_version":1}