{"id":"508f4046-833f-405c-92c3-53a3673a1929","arxiv_id":"2505.05741","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A density-guided DETR variant improves tiny object detection by 3.3 AP on AI-TOD-V2 and 2.5 AP on VisDrone over the D-FINE baseline.","lead":"Dome-DETR is a transformer object detector that estimates where tiny objects are concentrated and then uses that density map to prune computation and to place object queries. It reports large gains on aerial benchmarks, +3.3 AP on AI-TOD-V2 and +2.5 AP on VisDrone, with modest added compute.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Density-mask gating can prune low-density isolated objects: Eqs. (5)-(8) keep only near-maximal density windows, so missed DeFE predictions are structurally excluded; the paper never measures object-level mask recall.","rationale":"The central claim is that density-oriented gating improves both accuracy and efficiency. The gating is only as good as DeFE, and the threshold rule in Eqs. (5)-(7) is explicitly designed to keep at least one region, not all objects. This asymmetry creates a clear failure mode: sparse objects coexisting with dense clusters are pruned before detection. I agree with the reader's weakest_assumption. The paper's own analyses (Figures 6-7, Table 7) correlate window and query counts with ground-truth counts but never check whether individual ground-truth objects fall inside the retained mask, so the key quantity is unmeasured. The proposed test directly measures that quantity and an oracle upper bound. If the oracle shows large AP gains, the method is leaving performance on the table; if recall is high, the concern is resolved. I would keep the CONDITIONAL verdict until this measurement is reported; no move to reject because the mechanism is clearly specified and the ablations show consistent gains. I also note the appendix's Table 8 lists some higher-AP methods on VisDrone, but that is a framing issue; the density-recall question is the more mechanism-level risk for the core claim.","tokens_in":19514,"tokens_out":7221,"duration_ms":80812,"concrete_test":"Compute object-level mask recall on AI-TOD-V2 val for Dome-DETR-L: for each ground-truth box, record whether its center falls in a retained window M_W=1 from Eq. (8). Report recall overall and stratified by ground-truth count per image (e.g., 1-10, 10-100, >100). Then replace M_b with an oracle mask that is the union of the predicted mask and ground-truth density blobs at every ground-truth center, keep everything else fixed, and re-evaluate AP on the test split. If recall is below ~0.95 in sparse strata, or if oracle AP rises by more than ~0.5 AP, the density gate is removing valid objects and the reported gains are not from faithful density estimation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DeFE's predicted density map is the single gate for both computation and query allocation. Eq. (5) thresholds D_pred at T_b, and Eqs. (6)-(7) choose T_b as the highest threshold that leaves at least one active pixel (T_init=0.05, Delta-T fixed). Consequently, in any image whose maximum predicted density exceeds T_init, every object with predicted density below T_init is discarded from MWAS windows (Eq. 8) and from PAQI's flexible query set (Algorithm 1, line 13). Isolated tiny objects in mixed sparse/dense scenes systematically have lower predicted density than clusters, so the mask can remove exactly the hard sparse objects the method is supposed to help. The paper reports no object-level recall of M_b or M_W, no stratification by scene density, and no failure analysis of DeFE. QAR (Table 7) only counts queries versus ground-truth objects, not whether masks cover ground-truth centers. Without this measurement, the +3.3/+2.5 AP gains cannot be attributed to density-oriented gating rather than to pruning easy background; the central mechanism may be accidentally discarding difficult positives.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Dome-DETR proposes a DETR-based tiny-object detector built on D-FINE, with three new components: a Density-Focal Extractor (DeFE) that predicts a density heatmap, Masked Window Attention Sparsification (MWAS) that prunes background windows using a thresholded density mask, and Progressive Adaptive Query Initialization (PAQI) that adaptively allocates object queries using the same density signal. The paper reports +3.3 AP on AI-TOD-V2 and +2.5 AP on VisDrone over D-FINE, claims state-of-the-art performance, and provides ablations and efficiency analyses. The code is publicly available.","tokens_in":19798,"tokens_out":7109,"duration_ms":76499,"significance":"If the reported gains hold, the paper is a useful empirical contribution to tiny-object detection: it shows that a lightweight density prior can jointly guide feature sparsification and query allocation, and it provides thorough experiments on two standard aerial benchmarks with clear component ablations. The authors also ship code and report a meaningful comparison against D-FINE, a strong recent baseline. However, the central efficiency-and-accuracy claim is currently undermined by an internal contradiction in the VisDrone state-of-the-art claim, an apparent inconsistency in the DeFE formulation, and the absence of object-level mask-recall analysis for the gating mechanism.","major_comments":[{"comment":"As written, Eq. (2) defines F_G as the global average pooled feature over all spatial positions, and Eq. (3) computes D_pred by upsampling a function of F_G. Since F_G is spatially constant, Eq. (3) yields a constant density map, which contradicts the spatially varying density heatmaps shown in Figures 3 and 8 and the entire downstream gating mechanism. Please clarify the actual computation: if the density head operates on the spatially varying F'_S (perhaps with the GAP branch as an auxiliary channel-gate), rewrite Eqs. (2)-(3) accordingly; if the equations are correct, the model cannot produce the claimed masks.","section":"Section 3.2, Eqs. (2)-(3)"},{"comment":"The abstract and introduction claim state-of-the-art performance on VisDrone, but the paper's own Appendix Table 8 reports Focus&Detect at 42.0 AP and YOLC at 39.6 AP on the same validation split, both above Dome-DETR-L's 39.0 AP in Table 2. This is an internal contradiction in a load-bearing claim. Please either restrict the claim to DETR-based methods or to the latency-constrained setting, or provide a comparison that justifies the headline statement.","section":"Abstract, Section 1, Table 2, Appendix Table 8"},{"comment":"The binary mask M_b is created by thresholding DeFE predictions, and both MWAS window pruning (Eq. 8) and PAQI flexible-query filtering (Algorithm 1, line 13) discard regions whose predicted density is at or below the threshold. The paper never measures the object-level recall of this mask, i.e., the fraction of ground-truth object centers covered by M_b or M_W, nor does it stratify failures by scene density. Because any object missed by DeFE is structurally removed from the feature-enhancement path and from the flexible query set, the reported AP gains cannot be causally attributed to density-oriented gating without this measurement. Please add mask-recall statistics and a failure analysis (e.g., recall at thresholds, recall by object size and by per-image object count).","section":"Section 3.3-3.4, Eqs. (5)-(8), Algorithm 1"},{"comment":"The GFLOPs comparison mixes static and data-dependent values: Dome-DETR and DQ-DETR are marked with a star for average GFLOPs, while D-FINE and other baselines report fixed single-image GFLOPs. Appendix Table 6 shows that in dense scenes Dome-DETR-L reaches 398.9 GFLOPs, which is 71.4 GFLOPs above D-FINE-L's 327.5, not 31.2 as suggested in Section 4.5. To support the 'low computational cost' claim, report worst-case or matched-protocol GFLOPs (or latency on identical hardware) for all compared methods, and state explicitly which numbers are averages and which are static.","section":"Section 4.5, Table 1, Appendix Table 6"},{"comment":"The ablation study is reported from a single training run per configuration, and the per-component AP differences are 0.9-1.2 AP. On aerial benchmarks with single-seed evaluation, such differences can be within run-to-run variance. Please report the mean and standard deviation over at least three seeds, or provide statistical significance information, so that the claimed contributions of DeFE, MWAS, and PAQI are credible.","section":"Section 4.6, Table 3"}],"minor_comments":[{"comment":"The first sentence says 'Table 2 presents our main results on the AI-TOD-V2 test split,' but the AI-TOD-V2 results are in Table 1; the VisDrone sentence correctly refers to Table 2.","section":"Section 4.4"},{"comment":"The hyperparameter paragraph says T_init=0.05 'achieves AP 32.6', but Table 4 reports 33.3 for that setting; the text should match the table.","section":"Section 4.6, Table 4"},{"comment":"The header 'MS-WAS' should be 'MWAS' for consistency with the text and Figure 3.","section":"Table 3"},{"comment":"The variable D in Eq. (14) is described as 'the predicted density of the final mask-filtered tokens,' but it is not clear whether T is per-query, per-window, or a single scalar per image; please define D and the scope of the IoU threshold computation.","section":"Algorithm 1, Eq. (14)"},{"comment":"The section titled 'Frequently Asked Questions During Review' is not a standard paper section; consider removing it or folding its content into the main text or a normal limitations/discussion section, especially the explicit references to reviewer suggestions.","section":"Appendix E"},{"comment":"The conclusion repeats the phrase 'state-of-the-art results on the AI-TOD-V2 and VisDrone datasets'; this should be revised in light of the contradiction with Appendix Table 8.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The appendix's FAQ section and the inclusion of Focus&Detect and YOLC in Table 8 suggest the authors are aware of the VisDrone comparison, yet the abstract still claims state-of-the-art performance there; this needs to be reconciled in revision. The DeFE equations (Eqs. 2-3) appear to be a typographical error rather than a fundamental design flaw, but as written they invalidate the method description and must be corrected. The baseline D-FINE is a prior paper by overlapping authors, which is not disqualifying but should be kept in mind when interpreting the reported gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a competent engineering paper, not a breakthrough. It takes D-FINE and adds a light density head (DeFE), windowed sparse attention gated by that density map (MWAS), and adaptive query initialization (PAQI). On AI-TOD-V2 and VisDrone they get +3.3 and +2.5 AP over the D-FINE baseline, and the three-component ablation in Table 3 shows each adds something. The new QAR metric is a reasonable check on query coverage. The paper is clearly written and the code is promised.\n\nThe load-bearing concern is the density mask itself. Equations (5)-(7) set T_b effectively as T_init unless no pixel exceeds it, then step down. With T_init=0.05, any object whose predicted density is below 0.05 vanishes from the MWAS windows and from PAQI's flexible queries. That is exactly the regime of isolated tiny objects in mixed scenes. The loss penalizes underestimation, but the paper never reports object-level recall of the mask, nor stratified results by scene density. Without that, I can't tell whether the gains come from focusing on true foreground or from pruning hard background/positives. The stress-test note is right, and the absence of mask recall is a real gap. They report QAR (queries vs GT count), but not whether masks cover GT centers.\n\nOther soft spots: the VisDrone \"state-of-the-art\" claim is undercut by the appendix's own Table 8, where Focus&Detect gets 42.0 AP at higher resolution and much higher latency. That's a different regime, but omitting it from the main table while claiming SOTA is overclaiming. No error bars or multiple seeds, so the 2-3 AP deltas are plausible but not tight. The GFLOPs comparison is also apples-to-oranges since static baselines are compared to dynamic average GFLOPs. That's minor but should be stated clearly.\n\nThe citation pattern is fine: D-FINE is the baseline and self-cited, but that's legitimate given it's the foundation. DQ-DETR is properly discussed.\n\nBottom line: the paper deserves a serious referee. The method is well-specified, the ablations are honest, and the gains are meaningful for drone/satellite detection. But the authors need to measure mask recall, report variance, and settle the SOTA claim before this is ready to be taken as state of the art.\n\nRecommendation: engage with it, but require the recall analysis.","headline":"Dome-DETR is a solid empirical step for tiny object detection built on D-FINE, with clean ablations, but its density-mask gating can silently drop isolated low-density objects and the SOTA framing outruns the evidence.","tokens_in":20326,"tokens_out":2262,"would_cite":true,"duration_ms":22004,"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":"Dome-DETR claims that a learned density map can simultaneously guide sparse attention and query allocation, pushing tiny-object AP to 34.6 on AI-TOD-V2 and 39.0 on VisDrone.","keywords":["tiny object detection","detection transformer","density estimation","sparse attention","adaptive queries","aerial imagery","AI-TOD-V2","VisDrone"],"falsifier":"On AI-TOD-V2, compute the recall of the DeFE binary mask: count the fraction of ground-truth object centers that fall inside any retained window. If a substantial share of isolated small objects sits in pruned windows yet the reported AP still holds, the paper's explanation would be suspect; alternatively, ablating DeFE at inference by thresholding the density map at a fixed high value and observing a large AP drop would confirm that the mask, not the decoder, carries the gain.","tokens_in":19353,"feed_emoji":"🎯","tokens_out":4761,"duration_ms":50377,"temperature":0.7,"pith_summary":"Dome-DETR sets out to show that tiny-object detection in aerial imagery can be improved by letting a learned density map steer where the model spends computation and where it places its object queries. The paper claims this density-oriented manipulation pushes a D-FINE backbone to state-of-the-art results on two aerial benchmarks, with +3.3 AP on AI-TOD-V2 and +2.5 AP on VisDrone, while adding only modest parameters and FLOPs. A sympathetic reader should care because the approach directly attacks two known weaknesses of DETR-style detectors for small objects: shallow high-resolution features are expensive and diluted by background, and fixed query counts are wrong for scenes that range from one object to over a thousand. The paper argues that the same predicted density map can solve both problems at once.","feed_headline":"Density maps steer DETR queries to tiny object gains","feed_subtitle":"A learned density heatmap prunes background and allocates queries, adding 3.3 AP on AI-TOD-V2 at low extra cost.","key_machinery":"The load-bearing object is the predicted density map $D_{\\text{pred}}$: DeFE turns the shallowest backbone feature map into a normalized heatmap using dilated depthwise convolutions and a density head, supervised by Density Recall Focal Loss. A thresholded binary mask $M_b$ derived from this map (Equations 5–7) chooses which windows MWAS keeps, and the same mask filters flexible queries in PAQI. MWAS then applies self-attention inside retained windows plus an axis-permuted second attention pass for cross-window communication; PAQI splits selected queries into a fixed core set and a density-filtered flexible set, then applies a density-dependent IoU threshold in NMS. The density map is the shared signal that ties feature sparsification to query allocation.","core_discovery":"The central claim is that a single Density-Focal Extractor (DeFE) can produce a reliable pixel-level density heatmap of foreground objects from shallow backbone features, and that this heatmap can be reused to (i) prune background tokens from windowed self-attention via Masked Window Attention Sparsification (MWAS), and (ii) modulate the number, placement, and suppression threshold of decoder queries via Progressive Adaptive Query Initialization (PAQI). On the paper's own terms, the discovery is that these two uses are compatible and complementary: the density map focuses computation on informative windows while also giving dense scenes more queries and sparse scenes fewer, raising AP on AI-TOD-V2 from 30.1 to 33.3 for the small variant and reaching 34.6 AP for the large variant. The gains are largest for very tiny and tiny objects, which is the regime the components were designed for.","pith_inferences":["Editorial inference — a consequence the paper leaves implicit is that the three components are only as good as the density map: failures of DeFE would cascade into both MWAS and PAQI, so density-map recall should be reported alongside AP in follow-up work.","Editorial inference — the density heatmap could also serve other tasks such as object counting, saliency-guided crop selection, or adaptive video-frame sampling; the paper does not explore these uses.","Editorial inference — a testable extension is to train DeFE with DRFL but replace it at inference with ground-truth density maps; the performance gap would isolate how much of the gain comes from density-estimation accuracy versus the downstream gating mechanics.","Editorial inference — if DeFE is reliable, this line suggests that explicit density estimation may be a general cheap prior for DETR decoders in any image where foreground is sparse, not just aerial imagery."],"forward_implications":["If the reported results hold, density-guided sparse attention can recover tiny-object accuracy lost by deep low-resolution features without paying global-attention cost on high-resolution maps.","Dynamic query allocation would remove the need to hand-tune query counts per dataset; the same model adapts from sparse scenes to scenes with over a thousand objects.","The computational cost of the detector scales with scene density rather than being fixed, so average GFLOPs understate cost in dense scenes and overstate it in sparse ones.","Because the components plug into a D-FINE-style encoder-decoder, the same density-guided manipulation could be transferred to other DETR variants with shallow features."],"supporting_citations":[{"why":"D-FINE, the base detector whose encoder-decoder and training setup Dome-DETR extends; the reported AP gains are measured against its variants.","marker":"[27]"},{"why":"DQ-DETR, the prior dynamic-query tiny-object detector that Dome-DETR compares against and whose fixed-query limitation motivates PAQI.","marker":"[12]"},{"why":"AI-TOD-V2, the benchmark whose tiny-object AP is the paper's primary result; its extreme object sizes define the problem.","marker":"[41]"},{"why":"VisDrone, the second benchmark; its density variation tests query adaptation.","marker":"[52]"},{"why":"Deformable DETR, the sparse-attention ancestor whose broad multi-scale attention the paper identifies as inefficient.","marker":"[53]"},{"why":"RT-DETR, the real-time DETR whose HGNetv2 backbone and inference-speed design the method inherits.","marker":"[51]"},{"why":"DINO-DETR, a query-selection baseline that motivates the progressive query initialization in PAQI.","marker":"[47]"},{"why":"DDQ-DETR, which increases query count to 900 with fixed NMS and is compared as a dense-query alternative.","marker":"[49]"}],"fun_headline_variants":["Density heatmap focuses DETR attention and queries on tiny targets","Tiny object DETR: density-aware feature pruning and query allocation","Dome-DETR uses density maps to cut redundancy and boost AP on small objects","Density-oriented query shifts add 3.3 AP for tiny object detection","Learned density masks guide sparse attention and adaptive queries in DETR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes the DeFE density prediction is reliable enough to gate both computation and queries: the binary mask built from it prunes any window with zero predicted density and discards candidate queries outside the mask, so objects the density estimator misses are structurally removed from the detection path.","fun_headline_variants_meta":{"raw":{"variants":["Density heatmap focuses DETR attention and queries on tiny targets","Tiny object DETR: density-aware feature pruning and query allocation","Dome-DETR uses density maps to cut redundancy and boost AP on small objects","Density-oriented query shifts add 3.3 AP for tiny object detection","Learned density masks guide sparse attention and adaptive queries in DETR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1520,"prompt_tokens":949,"completion_tokens":571,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":471}},"tokens_in":565,"tokens_out":571,"duration_ms":6586,"temperature":1.0,"reasoning_tokens":471,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:58:18.974420+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On AI-TOD-V2, compute the recall of the DeFE binary mask: count the fraction of ground-truth object centers that fall inside any retained window. If a substantial share of isolated small objects sits in pruned windows yet the reported AP still holds, the paper's explanation would be suspect; alternatively, ablating DeFE at inference by thresholding the density map at a fixed high value and observing a large AP drop would confirm that the mask, not the decoder, carries the gain.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"D-FINE, the base detector whose encoder-decoder and training setup Dome-DETR extends; the reported AP gains are measured against its variants."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"VisDrone, the second benchmark; its density variation tests query adaptation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Deformable DETR, the sparse-attention ancestor whose broad multi-scale attention the paper identifies as inefficient."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RT-DETR, the real-time DETR whose HGNetv2 backbone and inference-speed design the method inherits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DINO-DETR, a query-selection baseline that motivates the progressive query initialization in PAQI."}],"review_version":1}