{"id":"9f728faf-70c1-484e-a2ae-5169276e56d1","arxiv_id":"2507.14855","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding a Gaussian-box Gromov-Wasserstein loss and Bayes-risk-based refinement to DETR detectors improves their AP on COCO and leukocyte datasets while producing localization uncertainty estimates.","lead":"An object-detection framework adds uncertainty estimates to DETR-style models by treating bounding boxes as Gaussian distributions and using a Gromov-Wasserstein distance term in the loss. On COCO it reports 0.1 to 1.4 AP gains across several DETR variants and state-of-the-art scores on two blood-cell datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bayes Risk Eq. 5 assumes the predicted mean equals the ground truth; in general the L2 risk is trace(Sigma) + ||mu_P - R||^2, so the trace-based reweighting is not grounded and may be an uncalibrated heuristic.","rationale":"The reader's weakest assumption identifies the same core issue: the Bayes Risk derivation assumes predictions are centered on ground truth, which is not the operating regime of the model. I agree with that reading. The central empirical claim is an AP improvement, and the largest identified source of gain is BRRM (+0.8 AP in Table 3). Since BRRM's behavior depends on token-wise reweighting by trace(Sigma_P), the validity of Eq. 5 is load-bearing: if trace(Sigma_P) is merely a variance estimate that is not calibrated to actual localization error, the risk-aware interpretation collapses, and the gains are at best an unexplained heuristic. My proposed check directly measures whether trace(Sigma_P) tracks the true squared error on held-out data. The paper's ablations are useful and the gains are plausible, but the absence of code, seeds, and error bars means these empirical claims cannot be independently confirmed from the preprint. The reader's CONDITIONAL verdict remains appropriate: release code, provide multi-seed statistics, and either supply a calibration analysis or revise the theoretical claims. Thus I recommend no change to the verdict.","tokens_in":16996,"tokens_out":11650,"duration_ms":137596,"concrete_test":"Using the released code or checkpoints, run H-DETR+ours on COCO val and log for each matched prediction the predicted covariance trace t_i = tr(Sigma_P) and the squared localization error e_i = ||mu_P - R||^2. Compute the mean gap E[e] - E[t] and a decile plot of E[e | t] against t. If the gap is systematically positive or E[e | t] is not increasing in t, Eq. 5 is not the model's actual risk and the BRRM gains cannot be attributed to Bayes-Risk filtering.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main reported gain comes from the Bayes Risk Refinement Modules (Table 3: +0.8 AP), whose justification is entirely Eq. 5. Appendix A.5 derives Eq. 5 by assuming each predicted coordinate is drawn from N(c_x, sigma^2) with the mean equal to the ground-truth coordinate c_x, and then integrating over a U(0,1) prior. In the real detector, the predicted box is a function of the image and its mean mu_P is not the ground truth. Under the paper's own Gaussian model, the expected squared L2 loss is tr(Sigma_P) + ||mu_P - R||^2. The second term is generally nonzero and can dominate, so trace(Sigma_P) is not the Bayes Risk of the model's regression loss. Consequently, the normalized risk vector T used to reweight output embeddings, classification loss, and matching cost is not measuring the minimum achievable loss, and Algorithm 1's confidence intervals are not calibrated to the actual localization error unless a calibration argument is supplied. The paper provides only a qualitative heatmap (Figure 6) rather than a calibration curve. The manuscript also does not specify how the per-query covariance Sigma_P is produced from DETR features, so it is not clear whether T is actually input-dependent or an arbitrary auxiliary output.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an uncertainty-aware plug-in framework for DETR-based detectors. Bounding-box predictions are modeled as a 4D Gaussian with learnable diagonal covariance, and the Gromov-Wasserstein distance is added as a regression loss. A 'Bayes Risk' equal to the trace of the covariance (Eq. 5) is used to reweight decoder output embeddings, classification loss, and matching cost, and an algorithm based on 95% confidence intervals is proposed to output a localization-uncertainty score. Experiments on COCO val2017 report AP improvements of 0.1 to 1.4 points over three DETR variants, and the framework is reported to achieve state-of-the-art AP on LISC and WBCDD leukocyte detection. Ablations in Tables 3 to 5 isolate the contribution of the GW loss and each Bayes Risk refinement module.","tokens_in":17297,"tokens_out":12283,"duration_ms":145305,"significance":"If the central derivation were valid, the framework would be a broadly applicable, low-cost enhancement with a useful uncertainty output, and the cross-domain results on medical images would increase its practical value. The paper deserves credit for reporting ablations that separate the GW loss from the refinement modules, for including a computational-complexity analysis, and for posting a code page. However, the theoretical grounding of the main performance gain is compromised by the Bayes Risk derivation, and the uncertainty quantification is not calibrated in a quantitative sense. The empirical gains on some baselines are small enough that missing seeds and implementation details are consequential.","major_comments":[{"comment":"The Bayes Risk identity Risk*=trace(Sigma_P) is derived under the assumption that each predicted coordinate is centered at the ground-truth coordinate: Section 3.1 sets mu_i = {c_x, c_y, w, h}, and Appendix A.5 uses the conditional model p(x|c_x)=N(c_x, sigma^2). In the actual detector, the predictive distribution is centered at the model output \\hat{R}, not at the ground truth R. For the paper's own Gaussian model, E[||\\hat{R} - R||_2^2] = trace(Sigma_P) + ||mu_P - R||_2^2, so trace(Sigma_P) is not the Bayes Risk of the L2 regression loss whenever mu_P differs from R. Since the normalized risk vector T in Eq. (7) is the sole input to the reweighting operations in Eqs. (8), (10), and (11), and Table 3 attributes +0.8 AP to BRRM, the central performance claim currently rests on an unsupported identity. The derivation must be replaced with one that integrates over the actual predictive distribution (with mean \\hat{R}) and a posterior over R, or the reweighting must be presented as an empirical heuristic backed by calibration evidence.","section":"3.3 and A.5 (Eq. 5)"},{"comment":"Algorithm 1 is underspecified: after dividing each of the four 95% confidence intervals into k bins, it says 'Form k bounding boxes R_i = (c_x_i, c_y_i, w_i, h_i)' without specifying how the bins are paired; if each interval is independently divided there are k^4 possible boxes, and choosing only the diagonal pairing is an unstated modeling decision that changes the output uncertainty. Moreover, the claimed reliability of the uncertainty estimate is supported only by the qualitative heatmap in Figure 6; no calibration curve, coverage rate, or correlation statistic is provided. Because the confidence intervals are built from the same sigma values whose meaning is called into question by the Bayes Risk issue, the uncertainty contribution needs either a quantitative calibration study or a clearly heuristic framing.","section":"Algorithm 1 and Section 5.3"},{"comment":"The manuscript never states how the per-query covariance Sigma_P is produced. Section 3.1 only says sigma_i is a learnable parameter with 0<sigma_i^2<=1, while Eq. (7) defines a per-query vector T and Eqs. (8) to (11) apply it per query. If the variances are not input-dependent outputs of a network head, the refinement modules cannot actually be query-adaptive; if they are, the architecture, initialization, and constraints need to be specified. This is a reproducibility gap for the central method.","section":"3.1 and 4.2"},{"comment":"All experiments appear to be single runs with no seeds reported. For Relation-DETR the reported improvements are +0.2 AP and +0.1 AP, which are within the usual run-to-run variation of 12-epoch COCO training, so the claim that the method 'consistently improves' all evaluated DETR variants is not statistically supported. At minimum, two or three seeds with means and standard deviations should be reported for the main comparisons, or released checkpoints should allow independent verification.","section":"Table 1"}],"minor_comments":[{"comment":"Equation (11) uses Risk* in the exponent of u but Risk*/4 in the exponent of s and in Eq. (7); this normalization inconsistency is likely a typo and should be corrected.","section":"4.2 (Eq. 11)"},{"comment":"The derivation assumes each ground-truth coordinate has a U(0,1) prior without stating this assumption in Section 3.3; the prior is also dimensionally unusual for width and height coordinates and should be stated explicitly when the Bayes Risk is introduced.","section":"Appendix A.5"},{"comment":"References [11] and [12] are the same paper (He et al., 2019) and should be merged.","section":"References [11] and [12]"},{"comment":"The caption notes that Relation-DETR uses its default classification loss without the Bayes Risk modification, but Section 5.2 does not discuss the implication; the text should be explicit that for Relation-DETR the complete proposed method was not applied.","section":"Table 1 caption"},{"comment":"Figure 6 is described as a heatmap but no colorbar or axis labels are shown, making it hard to read the claimed inverse relationship.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is a candidate for major revision rather than rejection: the empirical framework may be salvageable with a reframed uncertainty justification and calibration experiments. I would ask the authors to provide the implementation of the covariance head and to release code and checkpoints; given the small gains over Relation-DETR, single-run numbers will not be persuasive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the empirical work is decent and the ablations are honest; the theoretical framing around Bayes Risk doesn't hold up, so referee it but expect a rewrite.\n\nWhat's new: the specific combination of learnable-covariance Gaussian bbox modeling, a Gromov-Wasserstein loss, and Bayes-Risk-based reweighting of DETR internals. The paper reports AP gains on three DETR variants — +1.4 on H-DETR, +1.2 on DINO, and +0.1/+0.2 on Relation-DETR — and SOTA on LISC/WBCDD leukocyte detection. The ablations (Tables 3-5) isolate each component cleanly, and the Relation-DETR gains are honestly reported as small.\n\nThe soft spots are real. The Bayes Risk derivation in Appendix A.5 assumes each predicted coordinate is centered exactly on the ground truth and that ground-truth coordinates have uniform U(0,1) priors. In the actual detector, the predicted mean is not the ground truth, so the expected L2 loss is trace(Sigma_P) + ||mu_P - R||^2. Eq. 5 is therefore not a lower bound. This is load-bearing because the BRRM modules account for most of the gain (+0.8 AP in Table 3), and their only stated justification is Eq. 5. It may still work as a heuristic, but the paper should say so.\n\nThe Gromov-Wasserstein formula in A.3 doesn't contain means at all, which contradicts the motivating example in Figure 1. The formula looks misquoted from Delon et al.; at minimum, the authors need to reconcile it. The uncertainty quantification algorithm (Alg. 1) is heuristic, and the only validation is the heatmap in Figure 6, not a calibration curve.\n\nThe paper is not circular in the dangerous sense: the COCO and leukocyte numbers are against external baselines, and the ablations are coherent. But there are no seeds or error bars, and the code page is mentioned without the code itself being available.\n\nWho is this for? DETR practitioners who want a plug-in uncertainty-aware loss, and medical-imaging developers who need a localization-confidence score. A serious referee should see it because the empirical claims are reproducible enough to check and the idea is plausible. I would send it to review with major-revision recommendations: fix the theory, add calibration and multi-seed results, and actually release code.","headline":"Solid DETR plug-in with honest ablations, but the theoretical core—Bayes Risk as trace of covariance—rests on an assumption the model doesn't satisfy; referee it, but expect the theory to be rewritten.","tokens_in":17872,"tokens_out":4811,"would_cite":true,"duration_ms":50514,"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":"Modeling predicted bounding boxes as 4D Gaussians with a Gromov-Wasserstein loss improves DETR detectors and yields per-box localization uncertainty.","keywords":["object detection","DETR","uncertainty estimation","Gromov-Wasserstein distance","Gaussian bounding box","Bayes risk","bounding box regression","leukocyte detection"],"falsifier":"On a held-out split, compute for every predicted box the empirical squared error per coordinate, average it, and compare it with the average predicted variance; if the average predicted variance does not bound or track the empirical error, or if the off-diagonal covariances are large, the central Bayes Risk identity and the uncertainty score are not measuring what the paper claims.","tokens_in":16763,"feed_emoji":"🎯","tokens_out":6257,"duration_ms":64665,"temperature":0.7,"pith_summary":"This paper proposes a plug-in enhancement for DETR-style object detectors that treats each predicted bounding box not as four fixed numbers but as a four-dimensional Gaussian distribution with a learnable covariance matrix. The ground-truth box is modeled as a two-dimensional Gaussian derived from its inscribed ellipse, and the gap between the two distributions is minimized with the Gromov-Wasserstein distance rather than an IoU-style geometric term. From that probabilistic model the authors derive a Bayes Risk for $L_2$ regression that equals the trace of the predicted covariance, and they use it to reweight the decoder output embeddings, the classification loss, and the Hungarian matching cost. They also give an algorithm that converts the predicted covariance into a single localization-uncertainty score per box. On COCO the framework adds 0.1 to 1.4 AP across three DETR variants, and on the LISC and WBCDD leukocyte datasets it reports the highest AP among the compared detectors.","feed_headline":"Gaussian box model lifts DETR accuracy, adds uncertainty","feed_subtitle":"A plug-in loss and Bayes-risk reweighting add up to 1.4 AP on COCO and top leukocyte scores.","key_machinery":"The central object is the predicted 4D Gaussian $N_P(\\boldsymbol\\mu_P, \\boldsymbol\\Sigma_P)$ with learnable diagonal covariance, whose four coordinates are the box center, width, and height. The Gromov-Wasserstein distance between this 4D Gaussian and the 2D ground-truth Gaussian $N_g(\\boldsymbol\\mu_g, \\boldsymbol\\Sigma_g)$ provides a distributional regression loss; it is the mechanism that makes the covariance matrix trainable. The derived Bayes Risk, equal to the trace of $\\boldsymbol\\Sigma_P$, is the quantity that carries the uncertainty information into the three refinement modules. A confidence-interval algorithm over the predicted Gaussian converts the covariance into a single localization-uncertainty number per box.","core_discovery":"The paper's central claim is that explicitly modeling bounding-box prediction uncertainty, rather than treating predictions as point values, is a generally useful enhancement for DETR-based detectors. Concretely, the authors claim that replacing part of the deterministic regression objective with a Gromov-Wasserstein distance between a 2D ground-truth Gaussian and a 4D predicted Gaussian, and then using the derived Bayes Risk $\\mathrm{Risk}^* = \\sigma^2_{\\hat c_x} + \\sigma^2_{\\hat c_y} + \\sigma^2_{\\hat w} + \\sigma^2_{\\hat h}$ to refine output embeddings, classification loss, and matching cost, improves detection accuracy while also supplying an uncertainty estimate for each predicted box. The reported evidence is a consistent AP gain on COCO over H-DETR, DINO-DETR, and Relation-DETR, plus best-reported AP on the LISC and WBCDD leukocyte datasets. The authors further show that the uncertainty score inversely tracks a combined classification-IoU quality metric.","pith_inferences":["If the learned covariance truly tracks regression error, the same Bayes Risk scalar could serve as a query-ordering signal for DETR decoders beyond the three modules modified here, for example in iterative refinement stopping or active learning.","The trace identity depends on treating the four coordinates as independent and centered on the ground truth; testing whether off-diagonal covariance or systematic bias improves the uncertainty score would clarify whether the diagonal assumption is a real limitation.","The confidence-interval algorithm's division count $k$ trades inference time against uncertainty stability; a closed-form expected IoU under the predicted Gaussian might achieve similar calibration without sampling.","A direct comparison of predicted covariance with empirical error on held-out data would determine whether the reported uncertainty is aleatoric, epistemic, or a mix; the paper does not separate the two."],"forward_implications":["DETR variants that add the Gaussian modeling, Gromov-Wasserstein loss, and Bayes Risk refinement should gain accuracy without architectural changes; the paper reports +0.1 to +1.4 AP on COCO val2017 over three variants.","The same plug-in transfers to domain-specific detection: the enhanced H-DETR reports +1.4 AP on LISC and +1.9 AP on WBCDD over the prior best leukocyte detector.","The framework supplies a per-box localization uncertainty that is low for high-quality predictions and high for poor ones, matching the reported inverse relationship to the combined metric $s \\cdot u^{0.5}$.","The components are additive: the ablations attribute part of the gain to the Bayes Risk refinement modules, with the Gromov-Wasserstein loss as a prerequisite, and the remaining gains split between output embedding, matching cost, and classification loss.","The method also extends beyond DETR to one-stage detectors, with a VOC experiment showing YOLOv5 improved by 0.9 AP when the output embedding and classification loss are refined."],"supporting_citations":[{"why":"Introduces the DETR end-to-end transformer architecture that the proposed framework takes as its base.","marker":"[2]"},{"why":"Supplies the definition and analytical solution of the Gromov-Wasserstein distance between Gaussian distributions used as the regression loss.","marker":"[6]"},{"why":"The hybrid-matching H-DETR baseline whose output embedding, classification loss, and matching cost are refined and ablated.","marker":"[14]"},{"why":"The DINO-DETR baseline showing the enhancement transfers to a query-denoising DETR variant.","marker":"[38]"},{"why":"The Relation-DETR baseline, the strongest detector tested, where the method still reports a small AP gain.","marker":"[13]"},{"why":"Provides the IoU-aware classification loss and multiplicative matching cost that the Bayes Risk refinement extends.","marker":"[23]"},{"why":"Models bounding boxes as 2D Gaussians; the paper builds on this by making the prediction covariance learnable and four-dimensional.","marker":"[35]"},{"why":"The previous best leukocyte detector on LISC and WBCDD that the enhanced model is compared against.","marker":"[3]"}],"fun_headline_variants":["Uncertainty-aware DETR: plug-in gains AP and confidence","Gaussian boxes: DETR upgrade with built-in uncertainty","DETR gains AP via uncertainty modeling and Bayes risk","Uncertainty-aware DETR: +1.4 AP on COCO, top leukocyte"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that each predicted box coordinate is an independent Gaussian error centered on the true coordinate, with each true coordinate uniform on the unit interval; only under that setup does the Bayes Risk reduce to the trace of the predicted covariance, so the refinement and uncertainty scores inherit that assumption.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty-aware DETR: plug-in gains AP and confidence","Gaussian boxes: DETR upgrade with built-in uncertainty","DETR gains AP via uncertainty modeling and Bayes risk","Uncertainty-aware DETR: +1.4 AP on COCO, top leukocyte"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000649,"raw_usage":{"total_tokens":2988,"prompt_tokens":967,"completion_tokens":2021,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":1941}},"tokens_in":583,"tokens_out":2021,"duration_ms":14684,"temperature":1.0,"reasoning_tokens":1941,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:46:19.858116+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out split, compute for every predicted box the empirical squared error per coordinate, average it, and compare it with the average predicted variance; if the average predicted variance does not bound or track the empirical error, or if the off-diagonal covariances are large, the central Bayes Risk identity and the uncertainty score are not measuring what the paper claims.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the definition and analytical solution of the Gromov-Wasserstein distance between Gaussian distributions used as the regression loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The hybrid-matching H-DETR baseline whose output embedding, classification loss, and matching cost are refined and ablated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Relation-DETR baseline, the strongest detector tested, where the method still reports a small AP gain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The previous best leukocyte detector on LISC and WBCDD that the enhanced model is compared against."}],"review_version":1}