{"id":"116da418-7099-4b25-b201-5b1cb3920631","arxiv_id":"2504.16443","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"MGIoU averages 1D Generalized IoU values computed along shape normals, yielding a unified differentiable loss that outperforms task-specific losses on DOTA, Omni3D, ICDAR2017, and Waymo.","lead":"The authors propose MGIoU, a single loss function that measures overlap between convex shapes by projecting them onto their edge normals and averaging one-dimensional IoU scores. It is designed to replace task-specific losses in 2D/3D object detection, text detection, and trajectory prediction, with measured speedups of 10-40x.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MGIoU's claimed strong correlation with true IoU is never quantified, and its metric proof assumes a fixed projection set that the definition makes pair-dependent; a numerical check can settle both.","rationale":"The paper has genuine strengths: a novel projection-based loss, a released code URL, positive results on four benchmarks, and a correct derivation of the 1D GIoU simplification in Sec. 7. My concern is not that MGIoU is useless, but that two headline theoretical claims—'strongly correlated with IoU' and 'satisfies metric properties'—are not established by the manuscript. Defining A as the union of the two shapes' unique normals makes the identity converse and the triangle inequality depend on the pair being compared, and the appendix's triangle-inequality proof simply assumes a fixed A. The missing correlation measurement matters because the entire proposal is that averaging 1D GIoU over a handful of normals is a faithful proxy for high-dimensional overlap; if it is not, the 'unified objective' is just another heuristic loss, and the reported mAP gains could be dataset-specific. The proposed random-pair/triple numerical study would directly settle both points and can be run with the released code. Until then, conditional acceptance—keeping the paper but requiring this check—is the right posture, so the reader's CONDITIONAL verdict remains appropriate.","tokens_in":16562,"tokens_out":8509,"duration_ms":87684,"concrete_test":"Using the released code or a standalone implementation, sample 100k random pairs and triples of oriented rectangles and cuboids from the parameter ranges appearing in DOTA and Omni3D, with true IoU computed by polygon clipping or convex-hull intersection. Report (1) Pearson/Spearman correlation between MGIoU and true IoU, plus the fraction of pairs with MGIoU > 0.8 but true IoU < 0.3; (2) the maximum violation of LMGIoU(P,R) <= LMGIoU(P,Q) + LMGIoU(Q,R) over the sampled triples. Substantial violations or poor correlation would invalidate the 'strongly correlated' and 'metric' claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 defines the projection set A as the unique normals of both P and G, then asserts that LMGIoU satisfies non-negativity, identity, symmetry, triangle inequality, and scale invariance. Two load-bearing gaps appear in the appendix. First, Lemma 2's converse is asserted rather than derived: equality of all 1D GIoU values equal to 1 is taken to imply P=G 'given that A includes all unique normals,' but no argument or citation shows that interval equality on a finite set of edge/face normals determines a convex body, especially when A itself changes with the shapes' orientations. Second, the triangle-inequality proof in Sec. 8 explicitly assumes 'a fixed set of projection directions A for all pairs of shapes under consideration.' But for a triplet (P,Q,R), A(P,R), A(P,Q), and A(Q,R) are generally different—e.g., for rotated rectangles, A is aligned with each rectangle's own edges, so rotating one shape changes A. The proof therefore does not establish a metric. These are not cosmetic gaps: if the projection set discards information, MGIoU can be high when true IoU is low. For instance, two thin rectangles crossing at a small angle can have large interval overlaps along their own major axes while the true intersection area is tiny. The paper never reports any quantitative correlation between MGIoU and true IoU, despite the abstract's 'strongly correlated with IoU' claim. The experimental tables show aggregate mAP improvements, but those do not directly validate the correlation premise on the same shape pairs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MGIoU, a loss function for optimizing convex parametric shapes that projects both shapes onto a set of unique normals and averages one-dimensional GIoU values. Three variants are introduced: MGIoU for structured convex shapes, MGIoU+ for unstructured convex shapes via a convexity regularizer, and MGIoU− for minimizing overlap in trajectory prediction. The authors claim that MGIoU unifies position, size, and orientation optimization in a single differentiable objective, is strongly correlated with IoU, satisfies metric properties and scale invariance, and reduces loss-computation latency by 10–40x. Experiments are reported on DOTAv1.5 for oriented object detection, Omni3D for 3D 6-DoF detection, ICDAR2017 for quadrilateral detection, and Waymo for trajectory prediction, with code made publicly available.","tokens_in":16869,"tokens_out":3942,"duration_ms":39011,"significance":"If the claims are substantiated, MGIoU would be a practically valuable unification: one loss replacing task-specific IoU approximations across detection and trajectory tasks, with sizable speedups and consistent empirical gains on standard benchmarks. The empirical results are credible as reported, and the code availability strengthens reproducibility. However, the theoretical properties that the paper emphasizes—especially identity and the triangle inequality—rest on unproven or incorrectly scoped arguments, and the central 'strongly correlated with IoU' assertion is not quantified. These gaps do not necessarily invalidate the empirical contribution, but they need to be fixed or substantially qualified before the paper's stronger claims can be accepted.","major_comments":[{"comment":"The converse direction of Lemma 2 is asserted rather than proved: from equality of all 1D GIoU values (each equal to 1) the proof concludes P=G 'given that A includes all unique normals.' No argument or reference establishes that a finite set of edge/face normals determines a convex body, and the issue is especially delicate because A itself depends on the two shapes being compared. For rectangles and cuboids the claim may be true, but it must be demonstrated explicitly for the stated parametric families; without this, the identity property and the resulting metric claim are unproven.","section":"Appendix Sec. 8, Lemma 2"},{"comment":"The triangle-inequality proof explicitly assumes 'a fixed set of projection directions A for all pairs of shapes under consideration,' but the definition of MGIoU in Section 3.2 makes A depend on the pair (P,G). For a triplet of differently oriented rectangles, A(P,R), A(P,Q), and A(Q,R) are generally different. The averaging argument therefore does not apply to the actual MGIoU definition. Additionally, the assertion that 1−GIoU1D satisfies the triangle inequality and 'inherits' it from [27] is not demonstrated; the cited reference treats GIoU for boxes, not this particular interval distance. This is a load-bearing gap because the paper advertises metric properties as a theoretical guarantee.","section":"Appendix Sec. 8, Proposition 1 (Triangle Inequality)"},{"comment":"The abstract and introduction claim that MGIoU is 'strongly correlated with IoU,' but no experiment or theorem quantifies this correlation. Since MGIoU is an average of 1D GIoU values over a small set of normals, it can rank pairs differently from true IoU in simple cases, e.g., two thin rectangles crossing at a small angle can produce large interval overlaps along their own axes while the true intersection area is small. The paper's aggregate mAP improvements do not directly validate the correlation claim. Please add a controlled study on random valid shape pairs reporting Pearson/Spearman correlation between MGIoU (and its variants) and true IoU, with the shape families used in the experiments, or explicitly qualify the claim.","section":"Section 1 and Section 3.2"}],"minor_comments":[{"comment":"'Monocular 3D 6-DoF Object Regconition' contains a typo; it should be 'Recognition'.","section":"Section 4, heading"},{"comment":"The baseline name 'Distangled-L1+Chamfer' appears in the table and text; if this is intended to be 'Disentangled-L1+Chamfer', the spelling should be corrected for consistency with the cited work.","section":"Table 1 and Section 5.2"},{"comment":"The caption contains a formatting error: 'MGIoU+[A) unstructured quadrilaterals' should be 'MGIoU+ (A) unstructured quadrilaterals'.","section":"Figure 1 caption"},{"comment":"The statement that MGIoU latency is 'only slightly higher than L1' is misleading given the table shows 0.45 ms vs 0.03 ms (a 15x difference), though the absolute difference is small; please rephrase to avoid overstating speed, e.g., by noting both are below 0.5 ms.","section":"Section 5.1 and Table 2"},{"comment":"The Waymo experiment is trained on a 20% subset of the training set, so the absolute numbers in Table 3 are not directly comparable to full-benchmark results; this limitation should be stated more prominently in the main text, not only near the table.","section":"Section 4, trajectory prediction setting"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is real and the code availability is a plus, but the paper currently overclaims its theoretical properties. The metric proof and the identity lemma need to be fixed or explicitly downgraded to empirical observations, and the IoU-correlation claim needs quantitative support. The trajectory task result is based on a 20% training subset, which should be clarified to the editor when comparing with prior reported numbers. No concerns about novelty or scope; the work fits the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper introduces a new loss for convex shape regression: project both shapes onto their unique edge/face normals, compute 1D GIoU per direction, average. That is genuinely new relative to Gaussian approximations (KLD, GWD, KFIoU) and vertex-based losses (OKS, QRN). The experiments back the core claim: on DOTA, Omni3D, ICDAR2017, and Waymo, MGIoU variants beat task-specific baselines and cut loss latency by 10-40x. I believe the empirical story. The code is available, and the speedup is real.\n\nThe soft spots are all in the paper's stronger theoretical claims, and they are real. Section 8's triangle-inequality proof explicitly assumes a fixed projection set A for all shape pairs, but the definition of A in Section 3.1 is the union of the two shapes' unique normals, which changes with orientation. So the metric proof doesn't go through as written. The identity property (Lemma 2 converse) is asserted: equality of all 1D GIoUs is taken to imply P=G because A includes all unique normals, but no argument shows finite projections along pair-dependent normals determine a convex body. It may be true for same-family rectangles/cuboids; it is not proven for arbitrary orientations. The abstract also says MGIoU is 'strongly correlated with IoU,' but the paper never measures that correlation on shape pairs; the mAP gains are indirect. A simple numerical study would settle it. The stress-test scenario—two thin rectangles crossing at a small angle with large interval overlaps along their own major axes but tiny true intersection—is plausible and should be checked.\n\nMinor: experiments lack error bars, and the convexity weight lambda in MGIoU+ is not reported, so sensitivity is unknown.\n\nNone of this undermines the main empirical finding. The loss works and is fast, and the unification story is useful. But the advertised metric properties are currently unproven, and the 'strong correlation' claim is unsupported. Both are fixable in a revision: weaken the claims, or add a proof/correlation experiment.\n\nMy recommendation: send it to peer review. It deserves referee time. The reviewers should push on the theory and ask for correlation numbers, but the central idea and results are solid enough to warrant engagement.","headline":"A genuinely new, fast loss for convex shape regression with solid empirical gains, but the advertised metric properties are unproven and the IoU-correlation claim is unquantified.","tokens_in":17443,"tokens_out":2471,"would_cite":true,"duration_ms":25213,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a single loss, MGIoU, can replace domain-specific IoU approximations for every convex parametric shape by projecting shapes onto their unique normals and averaging one-dimensional GIoU scores.","keywords":["marginalized generalized IoU","convex shape optimization","oriented object detection","3D 6-DoF object estimation","quadrilateral detection","trajectory prediction","collision avoidance","differentiable loss"],"falsifier":"Take a large random sample of convex shape pairs, such as rotated rectangles or cuboids, compute true volume IoU and MGIoU for each pair, and check their rank correlation. The claim would be falsified if two non-identical shapes achieve identical projections on all unique normals, giving a perfect MGIoU score, while their true IoU is well below 1, or if the overall rank correlation between MGIoU and true IoU is weak on a large sample.","tokens_in":16372,"feed_emoji":"📐","tokens_out":8261,"duration_ms":77017,"temperature":0.7,"pith_summary":"The paper's central claim is that one differentiable loss function, MGIoU, can serve as a unified objective for aligning any convex parametric shapes, replacing the patchwork of task-specific IoU approximations used across 2D rotated detection, 3D 6-DoF pose estimation, quadrilateral detection, and trajectory prediction. The proposal is to project both shapes onto their unique edge or face normals, compute a one-dimensional GIoU along each normal, and average those values; for rectangles this needs only two projections and for cuboids three. The paper argues that this simple average stays strongly correlated with true IoU while making optimization stable and cheap, and that the same projection idea, when turned toward minimizing overlap, gives a collision-avoidance loss for trajectories. If the claim holds, one loss with no task-specific balancing could replace L1/L2, Chamfer, Gaussian-based IoU approximations, and keypoint-similarity losses while cutting loss computation latency by 10-40x.","feed_headline":"One loss replaces IoU approximations for any convex shape","feed_subtitle":"Project shapes onto their edge normals, average the 1D GIoUs, and one loss covers detection, pose, polygons, collisions.","key_machinery":"The object that carries the argument is the projection set $A$ of unique directional normals obtained from the faces or edges of the two shapes, together with the one-dimensional GIoU computed on each projection. For a rectangle, edge normals come in collinear pairs, so only two directions survive; for a cuboid only three, and for an ellipse the semi-axes serve as the normals. Algorithmically, each normal produces two intervals from the min and max of the projected vertices, and the 1D GIoU of those intervals is computed in closed form; the final score is the unweighted average over $A$. The same machinery is reused three ways: maximizing the average for structured shapes (MGIoU), adding a convexity regularizer on signed distances for unstructured polygons (MGIoU+), and taking the smallest 1D GIoU across normals with a softplus to minimize overlap (MGIoU-). This one mechanism is what lets a single loss address position, size, and orientation without balancing separate terms.","core_discovery":"The central claim is that generalized IoU can be marginalized over a small set of shape normals without losing what matters for optimization. Concretely, for a predicted shape P and ground truth G, MGIoU takes the union of their unique directional normals, projects both shapes onto each normal, computes the one-dimensional GIoU of those two intervals, and averages. For structured shapes this average is used directly in the loss $L=(1-\\mathrm{MGIoU})/2$; for unstructured shapes, a convexity regularizer is added, and for collision avoidance the smallest one-dimensional GIoU is penalized through a softplus. The paper claims this construction is symmetric, scale-invariant, zero only for identical shapes, and satisfies the triangle inequality, so the loss behaves like a metric and correlates with IoU. Empirically, the same machinery improves accuracy on oriented detection, 3D 6-DoF estimation, and quadrilateral detection, and reduces trajectory collisions by about 14 percent, while cutting loss computation latency by 10-40x compared with Gaussian and skew-IoU baselines.","pith_inferences":["Editorial inference: if the projection faithfulness holds beyond rectangles and cuboids, the same marginalization could give a differentiable IoU proxy for non-convex shapes by decomposing them into convex pieces; the paper does not claim this.","Editorial inference: a natural stress test is to compare MGIoU against true IoU on random convex polygons with many vertices; high rank correlation would confirm that the few unique normals are enough, while saturation would reveal the approximation's limit.","Editorial inference: the overlap-minimizing variant, which takes the smallest 1D GIoU across normals, resembles a Separating Axis Theorem certificate and could be transplanted to packing, layout, or collision-checking problems beyond trajectory prediction.","Editorial inference: one could replace Chamfer distance in shape registration with MGIoU+ and test whether overlap-based alignment improves on point-cloud fitting, since the loss is differentiable in vertex positions."],"forward_implications":["A single loss can replace task-specific objectives such as L1/L2, Chamfer, Gaussian Wasserstein, KLD, and OKS across 2D oriented detection, 3D 6-DoF estimation, and quadrilateral detection benchmarks, with gains in the reported metrics.","Because only a handful of normals are needed per shape, loss computation is 10-40x faster than KFIoU/GWD/KLD and close to L1 in latency.","MGIoU+ permits optimizing between shapes with different vertex counts or shape families, as long as the target has more vertices, opening polygonal fitting beyond quadrilaterals.","MGIoU- gives a differentiable overlap-minimization term that cuts collisions in multi-agent trajectory prediction by about 14 percent with no architectural change.","The metric properties make the loss scale-invariant and symmetric, so it does not need rescaling or task-specific tuning when input units change."],"supporting_citations":[{"why":"Defines GIoU and supplies the one-dimensional interval metric whose properties the triangle-inequality proof inherits.","marker":"[27]"},{"why":"Provides the 3D detection benchmark and the multi-term baseline loss that MGIoU replaces and improves.","marker":"[5]"},{"why":"Rotated-box baseline whose accuracy and latency MGIoU is compared against in the oriented detection evaluation.","marker":"[37]"},{"why":"Rotated-box baseline using KLD whose latency and accuracy are compared with MGIoU.","marker":"[38]"},{"why":"Rotated-box baseline using KFIoU, the slowest comparison point at about 51x MGIoU's latency.","marker":"[39]"},{"why":"Quadrilateral regression baseline compared on the ICDAR2017 dataset.","marker":"[12]"},{"why":"OKS-distance baseline compared on the ICDAR2017 dataset.","marker":"[19]"},{"why":"Trajectory prediction baseline whose architecture is kept unchanged to isolate the effect of MGIoU-.","marker":"[29]"},{"why":"Supplies the multi-agent trajectory benchmark and collision counts used in the MGIoU- evaluation.","marker":"[33]"}],"fun_headline_variants":["MGIoU: one loss, all convex shapes, 40x faster","Unify convex shape optimization with MGIoU loss","MGIoU: project normals, average 1D IoU, optimize all","MGIoU: metric-based, scale-invariant, and 40x speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction depends on the assumption that a shape's few unique edge or face normals carry enough geometric information that matching all one-dimensional projections is equivalent to matching the true volumes; if two distinct convex shapes had identical projections but different true IoUs, the loss would be blind to the difference.","fun_headline_variants_meta":{"raw":{"variants":["MGIoU: one loss, all convex shapes, 40x faster","Unify convex shape optimization with MGIoU loss","MGIoU: project normals, average 1D IoU, optimize all","MGIoU: metric-based, scale-invariant, and 40x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":2080,"prompt_tokens":1072,"completion_tokens":1008,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":688,"completion_tokens_details":{"reasoning_tokens":922}},"tokens_in":688,"tokens_out":1008,"duration_ms":9266,"temperature":1.0,"reasoning_tokens":922,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:03:39.420206+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a large random sample of convex shape pairs, such as rotated rectangles or cuboids, compute true volume IoU and MGIoU for each pair, and check their rank correlation. The claim would be falsified if two non-identical shapes achieve identical projections on all unique normals, giving a perfect MGIoU score, while their true IoU is well below 1, or if the overall rank correlation between MGIoU and true IoU is weak on a large sample.","supporting_citations":[{"cited_title":"Omni3D: A large benchmark and model for 3D object detection in the wild","cited_arxiv_id":null,"evidence_quote":"Provides the 3D detection benchmark and the multi-term baseline loss that MGIoU replaces and improves."},{"cited_title":"Learning high-precision bounding box for rotated object detection via kullback- leibler divergence","cited_arxiv_id":null,"evidence_quote":"Rotated-box baseline using KLD whose latency and accuracy are compared with MGIoU."},{"cited_title":"An end-to-end quadrilateral regression network for comic panel extraction","cited_arxiv_id":null,"evidence_quote":"Quadrilateral regression baseline compared on the ICDAR2017 dataset."},{"cited_title":"Yolo-pose: Enhancing yolo for multi person pose estimation using object keypoint similarity loss","cited_arxiv_id":null,"evidence_quote":"OKS-distance baseline compared on the ICDAR2017 dataset."},{"cited_title":"Motion transformer with global intention localization and lo- cal movement refinement","cited_arxiv_id":null,"evidence_quote":"Trajectory prediction baseline whose architecture is kept unchanged to isolate the effect of MGIoU-."},{"cited_title":"Scalability in perception for autonomous driving: Waymo open dataset","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-agent trajectory benchmark and collision counts used in the MGIoU- evaluation."}],"review_version":1}