{"id":"8b053795-116b-4fcd-9011-2db80b68c002","arxiv_id":"2505.21868","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cross-DINO improves small-object detection in DETR-like detectors by mixing MLP backbone features, a cross-coding fusion module, and a category-size soft-label loss, achieving 36.4% APs on COCO.","lead":"This paper combines a deep MLP backbone, a feature fusion module, and a size-aware loss to improve small-object detection in DETR-style detectors. It reports solid gains on COCO and several dedicated small-object benchmarks, making it relevant to anyone building detectors for drones, traffic, or dense scenes with tiny objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5)'s Boost loss may not implement the stated 'smaller objects get larger loss' mechanism, so the CS soft-label functional form is the least secure part of the +4.4 AP_S claim.","rationale":"","tokens_in":25908,"tokens_out":9154,"duration_ms":103336,"concrete_test":"On COCO, instrument Eq. (5) during the 12-epoch Cross-DINO-CLAP-Strip-T run and record the per-positive-sample loss term grouped by ground-truth size (small/medium/large) and confidence bin. If, under beta=1, the mean positive loss for small objects is not larger than for large objects, the mechanism in Sec. III-D is not realized. Then rerun the Table VI Boost-loss ablation replacing cs^beta with (a) a constant and (b) cs^(-beta), keeping the same beta-selection protocol. If AP_S changes by less than about 0.5 point, the CS functional form is not the active ingredient and the loss's contribution is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that Eq. (5) implements the stated 'Boost' mechanism. The text in Sec. III-D says that smaller objects are re-weighted through cs to increase their classification loss, but the positive term in Eq. (5) contains cs_i^beta as a multiplier. Since cs = sqrt((h/H)*(w/W)), cs^beta decreases as object size shrinks. With beta=1 (the COCO default), an 8x8 object in a 640x640 image gives cs=0.0125 and a combined positive factor of about 0.012, while an 80x80 object gives about 0.096, before the -log p term. The leading factor (1 - cshat^beta)^gamma does upweight small objects, but the cs^beta multiplier can cancel or reverse that effect, so the claimed monotone relation between size and loss is not guaranteed and may be inverted over part of the confidence range. Moreover, if cs is meant to be a soft label, Eq. (5) is not a cross-entropy against that label because the positive-class term -(1-cs)log(1-p) is missing; if cs is only a weight, the paper has not justified the square-root-area functional form. The only functional-form evidence is the beta sweep in Table X, which is on one dataset and does not compare other size transforms; there, beta=0.25 gives higher AP than the chosen beta=0.1, making the selection criterion unclear. Since Boost Loss contributes about +1.2 AP_S in the CLAP-Strip-T ablation (Table VI), the exact form of Eq. (5) is load-bearing for the headline +4.4 AP_S gain.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Cross-DINO extends the DINO DETR-style detector with three components aimed at small object detection: (1) CLAP-Strip-T, an adaptation of the Strip-MLP backbone to arbitrary input resolutions via overlapping crops; (2) a Cross Coding Twice Module (CCTM) that fuses backbone features with transformer encoder features; and (3) a Boost Loss with a new 'Category-Size' (CS) soft label that is meant to re-weight classification objective toward small objects. The paper reports consistent gains over DINO on COCO val2017 (36.4% AP_S vs. 32.0% with the CLAP-Strip-T backbone), WiderPerson, VisDrone2019, SODA-D, and smaller gains on AI-TOD, using a 12-epoch training schedule and comparable parameter counts.","tokens_in":26157,"tokens_out":7008,"duration_ms":68654,"significance":"If the claims hold, the paper would provide a useful practical recipe for improving small-object detection in DETR-like detectors, and it would demonstrate that MLP-style backbones can be adapted to dense prediction tasks without retraining. The paper includes a broad set of comparisons and ablations across five datasets, which is a strength. However, the central novelty that is most load-bearing—the Boost Loss and its CS soft-label formulation—is currently not adequately justified quantitatively, and the stated mechanism does not match the equation as written. The backbone and module contributions are plausible and well evaluated, but the loss contribution needs correction or re-analysis before the headline result can be fully credited.","major_comments":[{"comment":"The positive term of the Boost loss is α(1 - ĉs^β)^γ cs^β log(p_i). Since cs_i = sqrt((h_i/H)(w_i/W)) decreases with object size, the multiplier cs^β also decreases with decreasing size (for β=1, a 8×8 object in a 640×640 image gives cs≈0.0125 while an 80×80 object gives cs≈0.096). The claimed effect that 'smaller objects get larger loss' is therefore not implemented by this equation; the product (1 - ĉs^β)^γ cs^β is not monotone in size and can be inverted over plausible ranges. Table IX in the appendix computes the relative distance only for the (1 - ĉs^β)^γ factor, omitting the cs^β multiplier, so it does not support the stated behavior. Please provide a corrected loss form, a numerical demonstration that the full weight is monotone in the intended direction, or an explicit statement that the implemented loss is a different form than Eq. (5).","section":"Sec. III-D, Eq. (5)"},{"comment":"The paper repeatedly calls CS a 'soft label' and says it is used as 'new ground truth', but Eq. (5) is not cross-entropy with respect to the soft label cs. A standard soft-label cross-entropy for binary classification would include -(1-cs)log(1-p) in the positive-class term; here cs appears only as a multiplicative weight in the term -(cs^β)log(p). This is a semantic mismatch that affects the interpretation of the method. Please clarify whether cs is a soft label or a loss weight, and if it is a weight, justify the square-root-area functional form with a comparison to other size transforms (e.g., linear area, log-area, or relative-area-to-image).","section":"Sec. III-D, Eq. (4)-(5)"},{"comment":"The hyperparameter β is set differently across settings without a sensitivity study on COCO: β=1.0 for COCO-ResNet50, β=0.1 for COCO-Swin-T, β=0.1 for SODA-D, β=0.05 for AI-TOD. Table X, the only β ablation, is on VisDrone and shows that β=0.25 gives a higher AP (35.8) than the chosen β=0.1 (35.4), with only a small drop in AP_vt; the paper does not explain why β=0.1 is then preferred. Because Boost Loss is credited with gains up to +1.2 AP_S in the CLAP-Strip-T ablation (Table VI), the dependence of the method on a per-dataset, per-backbone β choice is load-bearing and must be analyzed on the main benchmark (COCO) and, ideally, with a principled selection rule.","section":"Table VI and Table X"},{"comment":"The CLAP method is essential for using Strip-MLP as a general backbone, but it is only validated through an ERF visualization (Fig. 11) and indirect detection results. No ablation compares CLAP with alternative resolution-adaptation strategies (e.g., tiling without overlap, padding to a multiple of the fixed size, or interpolation to the fixed size), and no sensitivity to the overlap sizes lw/lh is reported. A quantitative comparison or an ablation on the overlap parameter would significantly strengthen the claim that the proposed cropping scheme, rather than the backbone capacity difference, is responsible for the observed gains.","section":"Sec. III-B and Appendix 'Methodology of CLAP'"}],"minor_comments":[{"comment":"The notation for y_i is ambiguous: y_i could be a scalar class index or a one-hot vector, and the loss in Eq. (5) is written in a form that resembles binary classification. Please define the multi-class version (e.g., with sum over classes) and state how p_i is computed from the model's softmax.","section":"Eq. (4)-(5), notation"},{"comment":"The caption of Fig. 3 says that Cross-DINO detects 'harder' small objects with lower scores, resulting in a decrease in average confidence scores, which appears to contradict the motivation of Boost Loss as improving class prediction scores. The main text should discuss this trade-off explicitly.","section":"Fig. 3 and Sec. III-D"},{"comment":"The row 'DINO [49] Swin-T' cites reference [49] (detrex) rather than the original DINO paper; this should be clarified so readers know the baseline source.","section":"Table I"},{"comment":"The fixed weight of 2 for the encoder feature in Ecf = 2E·B'E' + B·(1-B'E') is introduced without an ablation or sensitivity analysis; a brief justification or a small study would make this design choice transparent.","section":"Sec. III-C, Eq. (3)"},{"comment":"In the caption of Table IX, γ is listed as 0.25, while the main text states the default γ=2.0; the inconsistency is confusing and should be corrected.","section":"Appendix, Table IX"},{"comment":"The paper says 'we introduce a new deep MLP model' but actually the deep MLP component is the existing Strip-MLP adapted by CLAP; rephrase to avoid overclaiming novelty of the backbone architecture itself.","section":"Sec. I and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is thorough in its empirical breadth, but the core defect in the Boost Loss equation is not a cosmetic issue: it changes the interpretation and, if the equation matches the code, likely inverts the stated size-dependent weighting. I would ask the authors to provide the actual loss implementation or a corrected derivation. The β selection without a COCO sensitivity study also weakens the claim of a general method. These are fixable in a major revision, but the authors need to substantiate the loss mechanism and re-run the affected ablations if the loss form is changed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the headline improvement is credible. The tables are internally consistent and the gains appear on multiple benchmarks. The CLAP adaptation of Strip-MLP to arbitrary input sizes is a neat, reusable idea, and the CCTM gated fusion is a reasonable way to push backbone detail into the encoder. The paper is also honest: it reports the AI-TOD regression, shows ablations for each component, and includes receptive-field and beta analyses in the appendix. Those are real strengths.\n\nThe soft spots are real too, and the most important one is in the Boost loss. The text says smaller objects get larger classification loss, but Eq. (5) multiplies the positive term by cs^beta, where cs is a size factor that shrinks as objects get smaller. With beta=0.1, the appendix's own numbers imply a 2x2 object gets a positive-term weight around 0.44 while an 8x8 object gets around 0.49, so larger objects get larger weights. The appendix's relative-distance analysis in Table IX only looks at the (1 - csh^beta)^gamma factor and ignores the cs^beta multiplier, so it does not support the stated mechanism. The beta ablation in Table X on VisDrone shows that beta=0.25 gives higher AP (35.8) than the chosen beta=0.1 (35.4), yet the paper uses beta=0.1 in its main results. That is an internal inconsistency, and it matters because Boost loss contributes roughly +1.2 AP_S in the CLAP-Strip-T ablation.\n\nThe other issues are more minor: no error bars or multi-seed runs, so gains of 0.2-0.4 AP may be within noise; beta is chosen per dataset with no sensitivity study on COCO; and the code URL is empty. There is also a small abstract/table mismatch about +1.1 AP versus +1.1 AP_S.\n\nThe paper deserves a serious referee. The system is promising, the experiments are broad, and the problems are fixable: re-derive or re-formulate the loss so it actually upweights small objects, justify or remove the cs^beta multiplier, run multi-seed ablations, and release the code. I would not desk-reject it, and if it lands in your pile, it is worth an hour of review time.","headline":"The +4.4 AP_S claim is probably real, but the Boost loss doesn't do what the paper says it does, and the beta selection is internally inconsistent.","tokens_in":26853,"tokens_out":5021,"would_cite":true,"duration_ms":46561,"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":"Cross-DINO claims that three coordinated fixes — a context-rich MLP backbone, a twice-gated encoder fusion, and a size-aware classification loss — raise COCO small-object AP_S from 32.0 to 36.4 while using fewer parameters than the DINO…","keywords":["small object detection","DETR-like detectors","deep MLP backbone","Cross Coding Twice Module","Boost loss","Category-Size soft label","transformer encoder","object detection"],"falsifier":"On COCO val2017 with the same 12-epoch DINO recipe, replace the CS target $cs_i = \\sqrt{(h_i/H)(w_i/W)}\\, y_i$ with $cs_i = ((h_i/H)(w_i/W))^p\\, y_i$ for several exponents p (e.g., 0.25, 0.5, 1, 2) and for a size-only reweighting with no class term; if AP_S stays roughly flat across p and across the no-class variant, the specific square-root category-size form is not what carries the gain.","tokens_in":25554,"feed_emoji":"🎯","tokens_out":7764,"duration_ms":69861,"temperature":0.7,"pith_summary":"Cross-DINO is a proposal for making DETR-style transformer detectors substantially better at finding small objects. The paper argues that three structural weaknesses — context-poor backbone features, blur introduced by repeated encoder attention, and classification confidence that drops as objects shrink — can each be fixed, and that fixing all three compounds. Its recipe replaces or augments the CNN backbone with a deep MLP that captures short- and long-range context in one pass, feeds fine-grained backbone detail back into the encoder through a twice-gated Cross Coding Twice Module, and reweights the classification loss with a new soft label that mixes class identity with object size. On COCO val2017, the combined model reaches 36.4% $AP_S$ with 45M parameters under a 12-epoch schedule, 4.4 points above the DINO baseline, and the gains carry over to WiderPerson, VisDrone2019, and SODA-D.","feed_headline":"Small-object AP on COCO rises from 32.0 to 36.4","feed_subtitle":"A deep-MLP backbone, a twice-gated encoder fusion, and a size-aware loss beat the DINO baseline with fewer parameters.","key_machinery":"The load-bearing mechanism is the Category-Size (CS) soft label, $cs_i = \\sqrt{(h_i/H)(w_i/W)}\\, y_i$, which fuses each object's class label with the square root of its area ratio to the image, and the Boost loss built on it; this is the component that directly targets low class-prediction scores for small objects. Two supporting mechanisms carry the rest of the argument: the CLAP-Strip-MLP backbone, an attention-free MLP whose overlapping-patch cropping makes it resolution-adaptive while mixing short- and long-range context, and the Cross Coding Twice Module (CCTM), a twice-gated fusion that re-injects fine-grained backbone detail into the blur-prone transformer encoder features. Together they supply richer initial features, counteract encoder blur, and reweight classification loss by object size.","core_discovery":"On the paper's own terms, the central discovery is that the small-object deficit of DETR-like detectors is not one problem but three, and that a detector built to address all three at once beats the standard DINO baseline by +4.4% $AP_S$ (36.4% vs. 32.0%) on COCO val2017 using fewer parameters (45M vs. 47M) under the same 12-epoch training. The first fix is a deep MLP backbone, CLAP-Strip-MLP, whose strip-wise token mixing captures both local and global context; a cropping-with-overlap scheme lets this fixed-resolution MLP accept arbitrary image sizes. The second is CCTM, which uses two gated crossings to let transformer encoder features absorb fine-grained detail from backbone features, countering the blur that repeated attention layers impose on small objects. The third is Boost loss, whose Category-Size soft label $cs_i = \\sqrt{(h_i/H)(w_i/W)}\\, y_i$ reweights each positive classification term by object size, raising the training pressure on exactly the small objects whose confidence scores lag. The paper reports consistent $AP_S$ gains across ResNet50, Swin-T, and CLAP-Strip-T backbones, and improvements over DINO on WiderPerson, VisDrone2019, and SODA-D, while noting that on AI-TOD the CLAP-Strip-T backbone's fewer channels hurt until the input resolution is increased.","pith_inferences":["Editorial extension: the CS soft label is a geometric reweighting of standard classification loss; a natural test is whether the square-root area ratio is special or whether any monotone size function (e.g., linear or log area) gives the same gain.","Editorial extension: CCTM's twice-gated fusion could be dropped into other encoder-decoder dense-prediction models, such as segmentation or keypoint detectors, wherever encoder downsampling blurs small structures.","Editorial extension: the AI-TOD result suggests a boundary condition — an MLP backbone with 31.25% of ResNet's channels underserves very tiny objects — so the paper's recipe may need a channel-wider MLP or higher input resolution to keep its edge in extreme tiny-object regimes.","Editorial extension: because the paper does not isolate feature-resolution from context-aggregation effects, an alternative explanation for part of the gain is that the deeper MLP simply provides a higher-quality feature pyramid; comparing Cross-DINO against DINO with a high-resolution ResNet variant would separate the two."],"forward_implications":["A DETR-like detector can gain 4.4 points of small-object AP on COCO val2017 without adding parameters, since Cross-DINO with CLAP-Strip-T runs at 45M parameters versus DINO's 47M.","The CLAP cropping scheme converts fixed-resolution deep MLP backbones into general-purpose, arbitrary-resolution backbones, making them usable in dense prediction tasks beyond classification.","Injecting backbone detail into encoder features through twice-gated fusion is a workable antidote to encoder feature blur for small objects.","Boost loss raises classification attention on small objects: on all three backbones tested, replacing the standard focal loss with Boost loss improves AP_S, with the largest single-component gain on CLAP-Strip-T (+1.2 AP_S).","The gains transfer across scene types: +3.7 AP on VisDrone2019, +4.3 AP on SODA-D, and +0.86 AP on WiderPerson with the Swin-T backbone."],"supporting_citations":[{"why":"DINO is the baseline detector that Cross-DINO extends; supplies the architecture, training recipe, and the AP_S numbers being surpassed.","marker":"[68]"},{"why":"Strip-MLP is the attention-free token interaction model whose fixed-resolution weights are made arbitrary-resolution by the CLAP cropping scheme.","marker":"[5]"},{"why":"DETR defines the end-to-end transformer detector framework and the encoder-decoder structure that the paper diagnoses as weak for small objects.","marker":"[7]"},{"why":"Deformable DETR provides the deformable-attention decoder used for iterative box refinement and prediction.","marker":"[74]"},{"why":"COCO supplies the val2017 benchmark and the AP_S metric on which the headline 36.4% vs 32.0% comparison is measured.","marker":"[30]"},{"why":"Focal loss is the classification loss that Boost loss replaces; the paper uses it as the comparison baseline in ablations.","marker":"[29]"},{"why":"ConvNeXt V2 contributes the Global Response Normalization used inside CCTM for channel re-weighting.","marker":"[61]"},{"why":"CrossViT motivates the cross-attention gating idea that CCTM adapts for fusing backbone and encoder features.","marker":"[8]"}],"fun_headline_variants":["Cross-DINO lifts small-object AP from 32.0 to 36.4","Beats DINO on small objects with 45M params, +4.4 AP_S","Three fixes for small detection: MLP, gated fusion, size-aware loss","Cross-DINO outperforms DINO by 4.4 AP_S on COCO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The recipe depends on the assumption that multiplying a box's class label by the square root of its area ratio to the image, used as the training target, is what improves small-object detection, rather than just any stronger emphasis on small boxes.","fun_headline_variants_meta":{"raw":{"variants":["Cross-DINO lifts small-object AP from 32.0 to 36.4","Beats DINO on small objects with 45M params, +4.4 AP_S","Three fixes for small detection: MLP, gated fusion, size-aware loss","Cross-DINO outperforms DINO by 4.4 AP_S on COCO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1675,"prompt_tokens":1200,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":816,"completion_tokens_details":{"reasoning_tokens":383}},"tokens_in":816,"tokens_out":475,"duration_ms":4805,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:20:12.046934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On COCO val2017 with the same 12-epoch DINO recipe, replace the CS target $cs_i = \\sqrt{(h_i/H)(w_i/W)}\\, y_i$ with $cs_i = ((h_i/H)(w_i/W))^p\\, y_i$ for several exponents p (e.g., 0.25, 0.5, 1, 2) and for a size-only reweighting with no class term; if AP_S stays roughly flat across p and across the no-class variant, the specific square-root category-size form is not what carries the gain.","supporting_citations":[{"cited_title":"Dino: Detr with improved denoising anchor boxes for end-to-end object detection","cited_arxiv_id":null,"evidence_quote":"DINO is the baseline detector that Cross-DINO extends; supplies the architecture, training recipe, and the AP_S numbers being surpassed."},{"cited_title":"Strip-mlp: Efficient token interaction for vision mlp","cited_arxiv_id":null,"evidence_quote":"Strip-MLP is the attention-free token interaction model whose fixed-resolution weights are made arbitrary-resolution by the CLAP cropping scheme."},{"cited_title":"Deformable detr: Deformable transformers for end-to-end object detection","cited_arxiv_id":null,"evidence_quote":"Deformable DETR provides the deformable-attention decoder used for iterative box refinement and prediction."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"COCO supplies the val2017 benchmark and the AP_S metric on which the headline 36.4% vs 32.0% comparison is measured."},{"cited_title":"Focal loss for dense object detection","cited_arxiv_id":null,"evidence_quote":"Focal loss is the classification loss that Boost loss replaces; the paper uses it as the comparison baseline in ablations."},{"cited_title":"Convnext v2: Co-designing and scaling convnets with masked autoencoders","cited_arxiv_id":null,"evidence_quote":"ConvNeXt V2 contributes the Global Response Normalization used inside CCTM for channel re-weighting."}],"review_version":1}