{"id":"7fa8167f-53f0-4817-8842-2b0fa63d7462","arxiv_id":"2507.12420","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"InterpIoU replaces geometric penalty terms in IoU-based bounding box regression with an IoU term on an interpolated box, improving small object localization.","lead":"This paper proposes InterpIoU, a new loss for object detection that adds a term based on the IoU between an interpolated box and the ground truth, which keeps gradients flowing when the predicted and true boxes do not overlap. The authors report small but consistent gains over prior IoU-based losses on COCO, VisDrone, and PASCAL VOC, with the largest improvements on small objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Non-overlap-gradient evidence uses Adam, which cancels the (1−α)=0.02 scale; the key YOLOv8/SSD experiments use SGD, where the interpolation term is ~50× weaker than an un-scaled penalty.","rationale":"The reader's weakest assumption was that the (1−α)=0.02 gradient scale may be insufficient, supported only by a toy simulation with Adam. I agree that this is the most load-bearing point, but the argument is sharper than the reader stated: the simulation's use of Adam actively neutralizes the 0.02 factor, making the toy result unrepresentative of the paper's own main experiments. Ultralytics YOLOv8 and SSD default to SGD, where gradient magnitude directly controls the update size. Thus the claimed mechanism—meaningful gradients for non-overlapping boxes—is validated only under an optimizer whose normalization hides the issue, while the experiments claiming practical gains run under an optimizer where the mechanism is potentially 50× weaker than competing geometric penalties. This does not prove the method fails; the reported improvements could stem from a regularizing effect in overlapping cases rather than from the non-overlap gradient. But it means the central mechanistic claim is not yet established for the actual experimental setting. I therefore keep the reader's CONDITIONAL verdict unchanged, with the condition sharpened to require either an SGD-based toy validation or an explicit demonstration that the interpolation gradient contributes comparably to standard penalty gradients during YOLOv8 training. I also note the paper's own admission that D-InterpIoU underperforms on COCO and the absence of WIoU/EIoU baselines, which further support the need for verification, but the gradient-scale/optimizer mismatch is the single most load-bearing concern.","tokens_in":16718,"tokens_out":25527,"duration_ms":278886,"concrete_test":"Run the Sec. 3.3 toy regression (or a simpler 1D/2D non-overlap case) with SGD (lr=0.01, momentum=0.937) instead of Adam, comparing InterpIoU with α=0.98 against CIoU/GIoU and against InterpIoU with α=0.5. If the α=0.98 curve fails to reduce the primary IoU loss within the same step budget (e.g., 200 steps), the non-overlap gradient is not meaningful under the optimizer used for YOLOv8/SSD, and the central mechanism is unsupported. A complementary check: instrument YOLOv8 training to measure the mean gradient norm contributed by the interpolation term on non-overlapping samples during the first epochs; it should be comparable to the distance-penalty gradient of CIoU to sustain the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central methodological claim is that LIoU(Bint,Bgt) supplies meaningful gradients when Bpred and Bgt are disjoint. Equation (19) shows this gradient enters through ∂Bint/∂Bpred = (1−α); with α=0.98 the gradient is scaled by 0.02. The only direct validation of sufficiency is the toy simulation in Sec. 3.3, which uses Adam at lr 0.01. Adam's per-coordinate normalization (m/√v) makes the step size nearly independent of this constant factor, so the simulation cannot reveal whether a 0.02× gradient is adequate. The paper's main experiments, however, train SSD and YOLOv8 under the default Ultralytics SGD recipe (the paper states YOLOv8 is trained with the Ultralytics framework and gives no optimizer override; DINO uses AdamW). Under SGD, the effective update from the interpolation term is ≈0.02× an ordinary gradient, so non-overlapping predictions receive extremely weak pulls. If this is insufficient, the claimed advantage over geometric-penalty losses (which provide O(1) gradients) would disappear in precisely the setting where YOLOv8/SSD results are reported. The paper also relies on the empirical condition d_i < l^g_i in Sec. 3.1, but α=0.98 makes that condition easy to satisfy, so the fragility is the gradient scale, not the overlap guarantee.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InterpIoU, a bounding-box regression loss that adds to the standard IoU loss a second IoU term computed between the ground-truth box and an interpolated box B_int = (1-alpha)B_pred + alpha B_gt. With alpha close to 1, the interpolated box overlaps the target even when the prediction is disjoint, providing a nonzero gradient signal without explicit geometric penalty terms. The paper also proposes Dynamic InterpIoU (D-InterpIoU), which sets alpha = clamp(1 - IoU(B_pred,B_gt), alpha_low, alpha_high) to strengthen gradients in low-IoU regimes. The central claims are that the interpolation term alone supplies meaningful non-overlap gradients, that handcrafted geometric penalties are unnecessary and even harmful, and that the proposed losses consistently improve small-object detection on VOC, VisDrone, and COCO with SSD, YOLOv8, and DINO. The paper supports these claims with a gradient derivation, toy regression simulations, an analysis of the bounding-box enlargement problem, and benchmark experiments.","tokens_in":16988,"tokens_out":9641,"duration_ms":101655,"significance":"If the central claims hold, the paper offers a conceptually simple alternative to the increasingly complex geometric penalties in IoU-based losses, and the proposed loss would be attractive for small-object and dense-scene detection where penalty terms are known to be unstable. The core construction is easy to implement and the gradient analysis in Eqs. (23)-(26) is correct and clearly presented. The paper also makes a falsifiable claim: interpolation-based IoU supervision should suffice without center-distance, aspect-ratio, or angle terms. However, the current evidence for that claim is not fully load-bearing: the supporting simulation uses Adam, which masks the small (1-alpha) gradient scale that matters for the SGD-based main experiments, and the D-InterpIoU gradient is not derived for the case where alpha depends on the prediction. The benchmark gains are small and are reported without variance, and the dynamic variant has dataset-specific clamp ranges. With additional experiments and clarifications the contribution could be solid, but in its present form the strong conclusions outrun the evidence.","major_comments":[{"comment":"The simulation that supports the claim that the interpolation term provides sufficient non-overlap supervision and that geometric penalties are unnecessary is run with Adam at learning rate 0.01 (Sec. 3.3). Because Adam normalizes each coordinate update by its running gradient magnitude, the reported curves are nearly invariant to the (1-alpha)=0.02 factor in Eq. (19), whereas the YOLOv8 and SSD experiments in Sec. 4.1 use SGD-style training under the Ultralytics framework with no reported optimizer override. The simulation also reports average loss/penalty values rather than final localization error. Please add SGD versions of the simulation, report final center and size errors, and include a real-training ablation that removes or rescales the interpolation term to isolate its effective contribution; otherwise the claimed advantage over O(1) geometric penalties in the benchmark settings is not established.","section":"Sec. 3.3, Eq. (19), Sec. 4.1"},{"comment":"For D-InterpIoU, alpha_dyn in Eq. (20) is a function of IoU(B_pred, B_gt), so B_int depends on B_pred through both (1-alpha) and alpha itself. The gradient decomposition in Eqs. (18)-(19), which treats alpha as constant, is therefore incomplete if alpha is not detached during backpropagation; the missing term is (B_gt - B_pred) * d(alpha)/d(B_pred). The paper should state explicitly whether alpha is detached and, if not, give the D-InterpIoU gradient. This matters because the 'gradient boost zone' discussion and Fig. 4 treat alpha as an externally varied parameter, which does not reveal the behavior of the differentiable loss; if alpha is detached, the dynamic mechanism is purely a loss-landscape change and the claim of an adaptive gradient boost should be qualified.","section":"Sec. 3.2, Eqs. (18)-(20)"},{"comment":"D-InterpIoU introduces two free parameters per dataset: the clamps in Sec. 4.2 are [0.95,0.99] on VOC, [0.60,0.99] on VisDrone, and [0.90,0.99] on COCO, while the ablation in Table 4 uses [0.00,0.99]. These ranges are not justified by a principled selection rule, and the statement in Sec. 1 that D-InterpIoU 'does not introduce complexity or handcrafted components' is therefore too strong. Please provide a validation-based selection rule for the clamps, or report sensitivity of the final metrics to the clamp values, and temper the handcrafted-component claim accordingly.","section":"Sec. 4.2, Table 4"},{"comment":"The reported gains over the best baseline are small (e.g., Table 3: YOLOv8 mAP 49.4 vs 49.3 for SIoU; DINO 49.7 vs 49.6 for PIoU), and in the DINO/COCO row D-InterpIoU is below PIoU (49.4 vs 49.6). No standard deviations, multiple seeds, or significance tests are provided. Since the paper claims consistent superiority over state-of-the-art losses, please add repeated-run statistics or explicitly restrict the claim to the settings where the difference is above noise.","section":"Tables 1-3"}],"minor_comments":[{"comment":"The sentence 'this term pushes w and h to decrease' is imprecise: the second term in Eqs. (25)-(26) is positive when w or h increases, so it counteracts the first term but does not by itself produce a decrease unless the intersection term is appropriately signed. Please rephrase.","section":"Sec. 3.4.2, Eqs. (25)-(26)"},{"comment":"The axes and curves in Fig. 5 are not fully labeled; please specify which subplot shows which metric and which curve corresponds to which loss, and state whether the plotted values are loss values, penalty values, or final errors.","section":"Fig. 5"},{"comment":"The simulation setup does not state the coordinate normalization or the box parameterization used in Eq. (12); please clarify whether the interpolation is applied to absolute [x,y,w,h] coordinates or to network-encoded offsets, since the equivalence matters for the real detectors.","section":"Sec. 3.3"},{"comment":"For SSD, YOLOv8, and DINO, please specify whether the loss is computed on decoded bounding boxes or on the network's raw regression targets; linear interpolation in Eq. (12) may not commute with the anchor or query encoding used by these detectors.","section":"Sec. 4.1"},{"comment":"The lower bound in Eq. (17) should explicitly note that the max is taken only over axes with d_i > 0; as written, the role of the max is not fully explained for overlapping axes.","section":"Sec. 3.1, Eq. (17)"},{"comment":"The paper states that alpha = 0.98 is used unless otherwise specified, but D-InterpIoU uses different alpha ranges on every dataset; please make the relationship between the static alpha and the dynamic clamp ranges explicit in the experimental setup.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the paper is within scope, but the strong thesis that geometric penalties are unnecessary rests on a simulation that is not representative of the SGD-based main experiments, and the D-InterpIoU gradient analysis is incomplete. The authors should be asked to address the gradient-scale issue, the alpha-differentiability question, and the per-dataset tuning of clamps before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the loss is real, simple, and works a little. The interpolation term — IoU between an interpolated box and the ground truth — is genuinely new and does not reduce to any of the cited prior losses. The dynamic version is also new and the intuition behind it is clear. Gradient derivation is correct, and the benchmark coverage is decent: three datasets, three detectors, and consistent, if small, gains, with up to +0.8 AP on small objects in COCO. The paper also gives a clean demonstration of how geometric penalties can cause bounding box enlargement, which is a useful observation.\n\nThe main soft spot is the load-bearing evidence for the central claim that handcrafted geometric penalties are unnecessary. That claim rests on the toy simulation in Sec. 3.3, which uses Adam at lr 0.01. Adam normalizes the step by the gradient magnitude, so the (1−α)=0.02 scale from Eq. 19 is essentially canceled. But the main experiments — YOLOv8 and SSD — are trained with SGD, where that scale remains. So the simulation does not establish that a 0.02× gradient is sufficient in the setting where the paper actually evaluates. This is a real flaw, and the paper should either provide an SGD-based simulation or temper the claim.\n\nOther soft spots are milder. The dynamic alpha ranges are tuned per dataset (VOC [0.95,0.99], VisDrone [0.60,0.99], COCO [0.90,0.99]), which is a form of hyperparameter fitting. EIoU and WIoU are discussed in related work but missing from the benchmark tables. No error bars, no code, and the ablation uses YOLOv8s for 100 epochs while the main results use YOLOv8m for 500, so the transfer is not automatic.\n\nNone of this is fatal. The loss itself stands as a reasonable new variant in a crowded field, and the empirical gains, though small, are consistent across frameworks. The paper would improve with: an SGD-based simulation, the missing baselines, repeated-seed statistics, and language that stops short of declaring all geometric penalties obsolete.\n\nI would send this to peer review. It deserves a serious referee. The idea is distinct, the gradient math is right, and the experiments, while modest, are real work. With revision, it could be a useful contribution to the BBR loss literature.","headline":"A genuinely new interpolated-box IoU loss with small but consistent gains; the simulation evidence for the strong 'penalties are unnecessary' claim does not survive contact with the actual optimizer in the main experiments.","tokens_in":17603,"tokens_out":2377,"would_cite":true,"duration_ms":28476,"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":"Adding an interpolated-box IoU term to regression loss fixes vanishing gradients and lifts small-object detection.","keywords":["bounding box regression","IoU loss","interpolated boxes","gradient vanishing","small object detection","object detection","dynamic interpolation","bounding box enlargement"],"falsifier":"A controlled experiment that keeps the loss value identical but multiplies only the interpolation term's gradient by zero would settle whether the non-overlap signal is the mechanism: if AP on COCO and VisDrone is essentially unchanged, the second term is not doing the claimed work. A complementary check is to train with a small fixed $\\alpha$ (e.g., 0.5), which gives a much larger interpolation gradient but less overlap margin; if regression from non-overlapping boxes still stalls, the rescue of vanishing gradients is not happening in practice.","tokens_in":16478,"feed_emoji":"🎯","tokens_out":10189,"duration_ms":109149,"temperature":0.7,"pith_summary":"This paper claims that the vanishing-gradient problem in bounding box regression--the loss stops teaching the model when the predicted box and the ground truth do not overlap--can be solved without any handcrafted geometric penalty terms. Its proposed loss, InterpIoU, adds to the ordinary IoU loss a second IoU term computed between the ground truth and a box that is a convex interpolation of the prediction and the ground truth. Because the interpolated box overlaps the ground truth even when the prediction does not, gradients keep flowing; because both terms are pure IoU, the loss stays aligned with the evaluation metric and cannot reward box enlargement. Simulations in the paper lead the authors to conclude that IoU alone is an ideal regression target, and they report consistent gains over previous IoU-based losses on COCO, VisDrone, and PASCAL VOC, most notably for small objects.","feed_headline":"InterpIoU fixes vanishing gradients, boosts small-box hits","feed_subtitle":"Adding an interpolated-box IoU term keeps gradients alive when boxes miss, improving small-object accuracy on COCO, VisDrone, and VOC.","key_machinery":"The interpolated box $B_{\\mathrm{int}}=(1-\\alpha)B_{\\mathrm{pred}}+\\alpha B_{\\mathrm{gt}}$ is the load-bearing object: a convex combination of prediction and ground truth that is guaranteed, for $\\alpha$ above a derived lower bound, to intersect the ground truth along both axes even when the prediction is far away. The paper derives the overlap guarantee $\\alpha > \\max_i d_i/(d_i + l_i^g)$ along each axis, uses the empirical observation that assigned boxes satisfy $d_i < l_i^g$ to justify $\\alpha > 0.5$, and sets $\\alpha=0.98$. The loss is the sum of the ordinary IoU loss and the IoU loss of the interpolated box; backpropagation through the second term gives a gradient scaled by $(1-\\alpha)$. D-InterpIoU replaces fixed $\\alpha$ with a clamped function of $1-\\mathrm{IoU}$, boosting gradients when IoU is low.","core_discovery":"The central discovery is that the loss $L_{\\mathrm{InterpIoU}}(B_{\\mathrm{pred}},B_{\\mathrm{gt}})=L_{\\mathrm{IoU}}(B_{\\mathrm{pred}},B_{\\mathrm{gt}})+L_{\\mathrm{IoU}}(B_{\\mathrm{int}},B_{\\mathrm{gt}})$ with $B_{\\mathrm{int}}=(1-\\alpha)B_{\\mathrm{pred}}+\\alpha B_{\\mathrm{gt}}$, $\\alpha=0.98$, supplies informative gradients for non-overlapping boxes while remaining a pure function of IoU. The second term is non-zero when the first term is flat, so it rescues gradient flow; its gradient is scaled by $(1-\\alpha)$, which the paper argues is sufficient. Since no geometric penalty term such as center distance or aspect ratio enters the loss, the optimizer's objective never conflicts with maximizing IoU, which the paper identifies as the root cause of the bounding box enlargement phenomenon. The dynamic variant D-InterpIoU sets $\\alpha_{\\mathrm{dyn}}=\\mathrm{clamp}(1-\\mathrm{IoU}(B_{\\mathrm{pred}},B_{\\mathrm{gt}}),\\alpha_{\\mathrm{low}},\\alpha_{\\mathrm{high}})$, creating a stronger gradient in low-IoU regimes. The paper's empirical claim is that this loss family outperforms prior IoU-based losses across three benchmarks and multiple detectors, with the largest margins on small and crowded objects.","pith_inferences":["Editorial inference: the same interpolation construction could transfer to other overlap metrics with flat regions, such as rotated-box IoU or segmentation IoU, where vanishing gradients are also a known obstacle.","Editorial inference: because the interpolation gradient is scaled by $(1-\\alpha)=0.02$ at the chosen $\\alpha=0.98$, an untested extension is to anneal $\\alpha$ during training to expose a stronger non-overlap signal early, which would isolate whether the magnitude or the mere presence of that signal drives the gains.","Editorial inference: if pure-IoU interpolation is as sufficient as the paper claims, the loss family should show its largest relative gains on datasets with extreme aspect-ratio or density distributions, where geometric penalties are most misaligned; this prediction is not tested in the paper."],"forward_implications":["For a prediction with zero overlap, the second term yields a nonzero gradient toward the ground truth, so non-overlap no longer means no learning signal.","Because both terms of InterpIoU are IoU-based, the optimizer has no incentive to inflate the predicted box to satisfy a geometric penalty, which the paper argues removes the bounding box enlargement problem.","The loss depends on no shape-, size-, or distribution-specific handcrafted parameters, which the paper identifies as the reason it stays stable for small and densely packed objects.","D-InterpIoU strengthens gradients in the low-IoU regime, acting as an implicit curriculum that helps the model escape flat loss regions.","On COCO, VisDrone, and PASCAL VOC, the paper reports that InterpIoU and D-InterpIoU improve localization accuracy over previous IoU-based losses across several detectors, with the largest gains on small objects."],"supporting_citations":[{"why":"Supplies the original IoU loss formulation that InterpIoU extends, providing the base term of the proposed loss.","marker":"[32]"},{"why":"Introduces GIoU and the bounding box enlargement problem, which the paper uses as the key failure mode it avoids.","marker":"[26]"},{"why":"Provides CIoU, a geometric-penalty baseline, and the simulation setup for controlled regression comparisons.","marker":"[38]"},{"why":"Contributes the 'major cases' observation behind the $\\alpha>0.5$ overlap bound and the Wise-IoU baseline.","marker":"[29]"},{"why":"Provides PIoU, a recent penalty-based baseline, and the simulation protocol used in Section 3.3.","marker":"[21]"},{"why":"Defines the COCO benchmark and its small-object AP metric used for the main evaluation.","marker":"[20]"},{"why":"Provides the VisDrone dataset used to evaluate dense small-object detection.","marker":"[3]"},{"why":"Provides the PASCAL VOC benchmark used for the comparison experiments.","marker":"[9]"},{"why":"The transformer-based detector on which InterpIoU is evaluated on COCO.","marker":"[34]"},{"why":"The real-time detector implementation used for most benchmark comparisons.","marker":"[19]"}],"fun_headline_variants":["InterpIoU: Interpolation-based IoU loss outperforms on small objects","InterpIoU loss eliminates box enlargement, boosts small objects","Dynamic InterpIoU adapts interpolation to improve small-box accuracy","InterpIoU: No more vanishing gradients for box regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the interpolation term's gradient, scaled by $1-\\alpha=0.02$, is large enough to actually drive training for non-overlapping boxes; the paper supports this with a toy simulation (Adam, learning rate 0.01) rather than a convergence proof or an isolated full-training experiment.","fun_headline_variants_meta":{"raw":{"variants":["InterpIoU: Interpolation-based IoU loss outperforms on small objects","InterpIoU loss eliminates box enlargement, boosts small objects","Dynamic InterpIoU adapts interpolation to improve small-box accuracy","InterpIoU: No more vanishing gradients for box regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000392,"raw_usage":{"total_tokens":2142,"prompt_tokens":1109,"completion_tokens":1033,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":725,"completion_tokens_details":{"reasoning_tokens":957}},"tokens_in":725,"tokens_out":1033,"duration_ms":10478,"temperature":1.0,"reasoning_tokens":957,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:46:55.056115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that keeps the loss value identical but multiplies only the interpolation term's gradient by zero would settle whether the non-overlap signal is the mechanism: if AP on COCO and VisDrone is essentially unchanged, the second term is not doing the claimed work. A complementary check is to train with a small fixed $\\alpha$ (e.g., 0.5), which gives a much larger interpolation gradient but less overlap margin; if regression from non-overlapping boxes still stalls, the rescue of vanishing gradients is not happening in practice.","supporting_citations":[{"cited_title":"Unitbox: An advanced object detection net- work","cited_arxiv_id":null,"evidence_quote":"Supplies the original IoU loss formulation that InterpIoU extends, providing the base term of the proposed loss."},{"cited_title":"Enhancing ge- ometric factors in model learning and inference for object detection and instance segmentation","cited_arxiv_id":null,"evidence_quote":"Provides CIoU, a geometric-penalty baseline, and the simulation setup for controlled regression comparisons."},{"cited_title":"Powerful-iou: More straightforward and faster bounding box regression loss with a nonmonotonic focusing mechanism","cited_arxiv_id":null,"evidence_quote":"Provides PIoU, a recent penalty-based baseline, and the simulation protocol used in Section 3.3."},{"cited_title":"Visdrone-det2021: The vision meets drone object de- tection challenge results","cited_arxiv_id":null,"evidence_quote":"Provides the VisDrone dataset used to evaluate dense small-object detection."},{"cited_title":"Everingham, S","cited_arxiv_id":null,"evidence_quote":"Provides the PASCAL VOC benchmark used for the comparison experiments."},{"cited_title":"Ni, and Heung-Yeung Shum","cited_arxiv_id":null,"evidence_quote":"The transformer-based detector on which InterpIoU is evaluated on COCO."},{"cited_title":"Ultralytics yolov8, 2023","cited_arxiv_id":null,"evidence_quote":"The real-time detector implementation used for most benchmark comparisons."}],"review_version":1}