{"id":"07fbff3e-e290-4ada-ab87-2fbc5122ef7e","arxiv_id":"2502.01565","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"By regressing Gaussian Cholesky parameters directly, GauCho removes oriented-box angular boundary effects and matches or improves AP against OBB heads on DOTA, HRSC, and UCAS-AOD.","lead":"GauCho is a new regression head for oriented object detection that outputs the parameters of a 2D Gaussian (via Cholesky factorization) instead of rotated boxes, avoiding the angle-boundary instability that complicates training. On aerial datasets such as DOTA, HRSC, and UCAS-AOD, it matches or slightly beats standard oriented-box heads while also supporting ellipse-shaped outputs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The theoretical boundary-discontinuity claim is not established: continuity of C(θ) does not rule out the near-boundary local minimum shown in Fig. 5, and the paper never tests whether GWD/ProbIoU remain non-convex in Cholesky coordinates.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the paper does not prove that directly regressing Gaussian parameters removes the angular-boundary training difficulty. My analysis sharpens that concern: continuity of the covariance matrix with respect to θ is not enough, because the Gaussian loss is still a function of the predicted covariance, and the local minimum shown in Figure 5 could persist unless the loss is convex or otherwise unimodal over the full Cholesky parameter space. For KLD the convexity argument may rescue the claim, but the paper does not make it; for GWD and ProbIoU no analogous argument is supplied. The proposed synthetic optimization directly tests the mechanism and would settle whether the concern lands. The verdict remains CONDITIONAL, as the reader recommended, because the head itself is a sensible and reproducible contribution with correct linear algebra and a broad experimental sweep; the issue is the overreach in the theoretical claim, not a demonstrated failure of the method. I do not see evidence of circularity, p-hacking, or derivation gaps that would justify rejection, and the code link plus MMRotate-based experiments are positive signs. The condition should be that the authors either provide the convexity/unimodality analysis for each loss or run the controlled optimization and report multi-seed statistics before claiming that GauCho theoretically mitigates boundary discontinuity.","tokens_in":18044,"tokens_out":14205,"duration_ms":138597,"concrete_test":"Run a controlled optimization study on the exact Appendix 7.1 setup: for a fixed GT OBB (w, h, θ) = (3, 1, 89°) and separately for GWD, KLD, and ProbIoU, minimize the loss over free Cholesky parameters (α, β, γ) with the center fixed, initializing at the Cholesky factor of the covariance at θ = -90°; compare with the same optimization over the OBB parameters (w, h, θ). If the Cholesky optimization consistently reaches the global minimum while the OBB optimization remains stuck, the proposed mechanism is supported for that loss; if both get stuck, or if the Cholesky optimization also stalls at the boundary, then the local minimum is intrinsic to the loss in Gaussian space and the theoretical claim in Sections 3.1 and 7.1 must be weakened. This isolates the head parameterization from all architectural and anchoring confounds present in Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the assertion in the abstract and Section 3 that directly regressing Cholesky parameters 'theoretically mitigates' boundary discontinuity. The only mathematical evidence is that the covariance elements in Eq. (15) are continuous and 180°-periodic in θ, plus the loss curve in Appendix 7.1. Continuity of the target representation does not by itself eliminate a near-boundary local minimum when the loss is evaluated on predicted covariances. Indeed, the paper's own Appendix 7.1 labels the causal mechanism a hypothesis ('We hypothesize that...'), and Figure 5 plots the loss only along the 1D OBB angle curve, not over the full Cholesky output space. That curve cannot show whether an optimizer free to move in (α, β, γ) escapes the -90° basin. For the KLD loss in the usual predicted-to-target direction, the loss is convex in the predicted covariance, so the paper's claim may be salvageable for that loss; but no such argument is given for GWD or ProbIoU, and the paper claims compatibility with all three. The experimental section does not isolate the mechanism: Table 1 changes the head parameterization, anchor/stride scaling, and decoding simultaneously, and results are single-run with mixed outcomes, e.g., R3Det-KLD DOTA AP50 is 38.90 for OBB versus 37.65 for GauCho, while FCOS-KLD moves in the opposite direction. Thus the central causal claim is under-supported exactly where it is needed to explain why GauCho should reduce boundary-related training difficulty.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GauCho, a regression head for oriented object detection that directly outputs the lower-triangular Cholesky factors (α, β, γ) of a 2D Gaussian covariance matrix, instead of the usual oriented-bounding-box (OBB) parameters (x, y, w, h, θ). The authors derive bounds relating the Cholesky parameters to OBB dimensions, instantiate the head in anchor-free and anchor-based detectors, advocate oriented ellipses (OEs) as an alternative output representation, and evaluate the head with GWD, KLD, and ProbIoU losses on FCOS, RetinaNet, R3Det, and RoI-Transformer over HRSC, UCAS-AOD, and DOTA v1.0/v1.5. The paper's central theoretical claim is that this parameterization is continuous in orientation and therefore mitigates the angular boundary discontinuity problem.","tokens_in":18397,"tokens_out":4491,"duration_ms":38731,"significance":"If the continuity argument were sufficient to guarantee the mitigation claim, GauCho would be a simple, loss-agnostic representation change applicable to a broad family of existing detectors and Gaussian-based losses. The experimental breadth is a genuine strength: four detectors, three losses, and three datasets using standard MMRotate configurations, with competitive DOTA results when multiscale training is used. The OE analysis is also a useful contribution for thinking about circular and ill-oriented objects. However, the central 'theoretically mitigates' claim is weaker than stated: continuity of the target mapping does not by itself ensure the loss landscape has no problematic near-boundary local minima, and the paper's own appendix frames the causal mechanism as a hypothesis. The practical value of the head as an alternative representation is supported by the experiments, but the deeper theoretical claim needs either additional analysis or a more modest framing.","major_comments":[{"comment":"The central claim that directly regressing Cholesky parameters 'theoretically mitigates' the boundary discontinuity problem is not established. The argument shows only that the covariance elements a, b, c in Eq. (15) are continuous and 180°-periodic in θ; this does not imply that GWD, KLD, or ProbIoU, viewed as functions of the predicted Cholesky parameters (α, β, γ), have no problematic local minima near the angle boundary. Appendix 7.1 itself labels the mechanism as a hypothesis, and Figure 5 plots the loss only along the one-dimensional OBB angle curve, not over the full Cholesky output space. Please either provide a proof or a numerical optimization-landscape study for the relevant losses in (α, β, γ) coordinates, or weaken the claim to 'empirically improves orientation consistency'.","section":"Section 3.1 and Appendix 7.1"},{"comment":"Proposition 3.2 as stated is incorrect: the text claims |γ| < √λmax − √λmin, but the proof's Eq. (27) yields f(x*) = (√λw − √λh)^2, so the inequality is attained and should be ≤. This matters because Eq. (13) uses the bound as the default scale for γ; the construction remains valid with ≤, but the proposition must be corrected.","section":"Section 3.3.1 and Appendix 7.2"},{"comment":"The experimental evidence for the central mechanism is mixed and underpowered. On DOTA v1.0, GauCho improves FCOS across all losses, but for R3Det it decreases AP50 under KLD (38.90 → 37.65) and ProbIoU (38.91 → 37.89), and for RoI-Transformer all three losses give lower AP50 (e.g., KLD 45.96 → 44.32). Since all results appear to be single runs, seed variability cannot be assessed. To support the claim that GauCho mitigates the boundary problem rather than merely serving as an alternative parameterization, the key comparisons should include multiple seeds or error bars, and the head change should be isolated from the anchor/stride scaling and OBB-to-OE decoding changes.","section":"Table 1 and Section 4"}],"minor_comments":[{"comment":"There are several typos: 'ambuiguity' in Sections 1 and 2, 'Choleky' in Proposition 3.2, and 'Not that' in Section 5, which should be 'Note that'.","section":"Throughout"},{"comment":"The number of DOTA categories is inconsistent: Section 5 says 'nine of the 15 categories', while Appendix 7.3 says 'nine of the sixteen categories' and Figure 6 says 'sixteen'; DOTA v1.0 has 15 categories.","section":"Section 5 and Appendix 7.3"},{"comment":"Equation cross-references are inconsistent: Section 3.1 refers to Eq. (15) for the covariance matrix, but the equation appears as Eq. (3), and Section 3.3.1 refers to 'Eq. (1) to Eq. (17)', which does not match the displayed equation numbers.","section":"Section 3.1 and 3.3.1"},{"comment":"The column header 'HRSC (OBB)' is potentially confusing because the table also reports OE-based metrics for UCAS-AOD; a clearer caption or sub-headers would help.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"I see no citation or authorship concerns. The main issue is the gap between the theoretical claim and the evidence; the practical, empirical contribution is reasonably solid. If the authors can either supply landscape analysis for the losses in Cholesky coordinates or reframe the theoretical claim, the paper would be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about GauCho. First, the core idea is genuinely new: no one has regressed Cholesky parameters of a Gaussian covariance directly for oriented object detection, and it slots cleanly into existing detectors with Gaussian losses. Second, the paper's loudest claim — that it \"theoretically mitigates\" the boundary discontinuity problem — is a hypothesis the authors themselves label as such in Appendix 7.1. The math proves continuity of the covariance as a function of angle, but that does not rule out the near-boundary local minimum shown in Figure 5. A 1D loss curve along the OBB angle tells you nothing about whether an optimizer free in (α, β, γ) space escapes that basin. So the theoretical contribution is real but weaker than the abstract suggests.\n\nThe paper does several things well. Proposition 3.1 is immediate, and Proposition 3.2 is proved correctly (modulo a strict-vs-non-strict inequality slip in the text; the bound is attained). The anchor-free and anchor-based head formulations are sensible, and the bounds relating Cholesky entries to box dimensions are practically useful. The experiments sweep four detectors, three losses, and four datasets on the MMRotate configs — a broad, honest test. The OE discussion is thoughtful and the orientation-error analysis on HRSC is a nice addition.\n\nSoft spots, in proportion. The causal story for boundary discontinuity is under-supported, as stressed above. The empirical results are mixed: GauCho helps FCOS consistently, but for R3Det and RoI Transformer it often loses to the OBB head, and the tables contain single-run numbers with no error bars or seeds. The paper attributes the differences to \"random seed selection\" without providing evidence. Code is promised but not yet available, which matters for a representation claim.\n\nWho gets value: researchers working on rotated object detection, especially those using Gaussian losses, will find the representation and bounds useful. It deserves a serious referee — not a desk reject — but the referee should send it back for a major revision: tone down the theoretical claim to what is proven, add multi-seed statistics, and release code. I would not accept it as-is, but the underlying idea is sound and likely to be cited once it lands in usable form.","headline":"GauCho is a genuinely new regression head with correct math and broad experiments, but the boundary-discontinuity claim is a labeled hypothesis, not a proven theorem, so the paper needs revision and reproducibility artifacts before acceptance.","tokens_in":18905,"tokens_out":1349,"would_cite":true,"duration_ms":14715,"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":"Oriented object detection can sidestep angle discontinuities by regressing Gaussian Cholesky parameters directly instead of oriented bounding boxes.","keywords":["oriented object detection","Gaussian distribution","Cholesky decomposition","boundary discontinuity","oriented ellipses","remote sensing","DOTA","Gaussian loss"],"falsifier":"Train an otherwise identical detector to regress the Cholesky parameters (α, β, γ) directly from a Gaussian ground truth, using a Gaussian loss, and plot the loss as a function of predicted orientation for a fixed anisotropic object. If a second local minimum persists near the boundary (e.g., near -90 degrees for a target at 89 degrees) that is almost as deep as the global minimum, the causal claim that the mapping is the culprit would be refuted; if the landscape becomes single-minimum, the claim survives.","tokens_in":17854,"feed_emoji":"🛰️","tokens_out":5001,"duration_ms":39956,"temperature":0.7,"pith_summary":"Oriented object detection usually regresses rotated rectangles described by center, size, and angle, and the angle parameter jumps at the wrap-around boundary, confusing training. This paper proposes GauCho, a regression head that outputs the parameters of a 2D Gaussian distribution directly, specifically the lower-triangular Cholesky factors of its covariance matrix, instead of passing through oriented bounding boxes. Because the Cholesky parameters are continuous functions of rotation, the head theoretically removes the boundary discontinuity that plagues angle-based regression, while remaining compatible with existing Gaussian-based losses such as GWD, KLD, and ProbIoU. The paper also argues for reporting detections as oriented ellipses, which remove the meaningless orientation assigned to circular objects. On DOTA, the proposal matches or improves the standard OBB head, with the largest gains on the anchor-free detector FCOS.","feed_headline":"Regress Gaussians, not boxes: new head tames rotated-detection angles","feed_subtitle":"By regressing Cholesky parameters of a 2D Gaussian, oriented detectors stay angle-continuous and beat DOTA baselines.","key_machinery":"The Cholesky decomposition of the 2D covariance matrix, C = LL^T with L = [[α, 0], [γ, β]] and α, β > 0, γ ∈ R. This gives a unique, continuous parameterization of a Gaussian distribution: unlike the covariance entries (a, b, c), the parameters (α, β, γ) are unconstrained except for positivity of α and β, and unlike OBB parameters (w, h, θ) they are 180-degree periodic and continuous in rotation. The paper uses it as the regression target of detection heads, with multiplicative offsets relative to anchors or strides, and proves bounds relating α, β, γ to object dimensions so the head can be initialized from anchor geometry.","core_discovery":"The paper claims that the angular boundary problem in oriented object detection is caused by the mapping from oriented bounding boxes to Gaussians, not by Gaussian loss functions themselves, and that regressing Gaussian Cholesky parameters directly should therefore mitigate it. GauCho parameterizes a 2D Gaussian by its mean (x,y) and a covariance matrix C, whose positive-definiteness is enforced by writing C = LL^T with lower-triangular L, whose entries α, β, γ the network outputs. The paper proves bounds relating α, β, γ to the OBB dimensions w and h and shows a bijective continuous relation between the Cholesky parameters and the Gaussian, so any Gaussian-based regression loss can be applied without angle discontinuity. It also introduces oriented ellipses as a decoding target and shows that they are a natural output for Gaussian-based detectors, especially for near-circular objects.","pith_inferences":["A testable extension: if the causal claim is right, then any future improvement to Gaussian loss functions will inherit the boundary-discontinuity fix without extra machinery; one way to test this is to inject perfect Cholesky ground truth and check whether the two local minima in the loss landscape disappear.","A natural extension beyond 2D is to regress Cholesky factors of 3D covariance ellipsoids for oriented 3D detection, where the same continuity argument should hold.","The OE representation could become a practical output format for detectors on datasets with many circular instances, replacing arbitrary OBB angles with a representable circle; measuring IoU against masks would quantify the gain.","GauCho's bound equations suggest that anchor boxes could be designed directly in (α, β, γ) space; testing whether such anchors give better recall than OBB-space anchors would isolate the benefit of the parameterization."],"forward_implications":["Any Gaussian-based regression loss — GWD, KLD, or ProbIoU — can be trained with the GauCho head without modifying the loss, since the head outputs the same Gaussian parameters the loss consumes.","Anchor-free and anchor-based, one-stage and two-stage detectors can switch to GauCho by changing only the regression head, keeping the rest of the architecture intact.","GauCho reduces orientation errors: on HRSC the average orientation error drops from 1.36 degrees with the OBB head to 1.11 degrees, with smaller errors in every orientation bin.","On DOTA v1.0 with multiscale training, GauCho paired with RoI-Transformer reaches 80.61 AP50, better than the OBB-head methods compared.","Decoding detections as oriented ellipses raises AP75 on UCAS-AOD and gives a representation that is well defined for circular objects."],"supporting_citations":[{"why":"Defines the ProbIoU loss and the OBB-to-Gaussian mapping, and introduces the encoding ambiguity for circular objects that GauCho addresses with oriented ellipses.","marker":"[20]"},{"why":"The FCOS anchor-free detector that the GauCho head is built on and that shows the biggest DOTA gains.","marker":"[22]"},{"why":"The DOTA dataset used for the main experimental comparisons.","marker":"[24]"},{"why":"Recent work that notes Gaussian-based losses can still suffer boundary discontinuity, providing the motivation for the paper's causal hypothesis.","marker":"[27]"},{"why":"The GWD loss, one of the Gaussian-based regression losses tested with GauCho.","marker":"[32]"},{"why":"The KLD loss, another tested Gaussian-based loss, and the loss used in the paper's loss-landscape illustration.","marker":"[33]"},{"why":"The prior work that maps OBBs to Gaussian distributions and provides the anchor-based regression scheme that GauCho adapts.","marker":"[35]"},{"why":"Recent work on boundary discontinuity in angle regression that, together with [27], motivates the paper's claim about the residual discontinuity.","marker":"[38]"},{"why":"The RoI-Transformer two-stage detector that, paired with GauCho, yields the best DOTA result.","marker":"[2]"}],"fun_headline_variants":["Cholesky Gaussian head removes angular boundaries in oriented detection","GauCho: direct Gaussian regression for oriented objects, no boundary","Boundary-free: new head outputs Cholesky Gaussians for rotation","Angle continuity from Cholesky: GauCho beats DOTA baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The advantage of GauCho rests on the hypothesis that the angular boundary problem observed with Gaussian losses is caused by the OBB-to-Gaussian mapping, not by the Gaussian loss function itself; the paper illustrates but does not prove this causal chain.","fun_headline_variants_meta":{"raw":{"variants":["Cholesky Gaussian head removes angular boundaries in oriented detection","GauCho: direct Gaussian regression for oriented objects, no boundary","Boundary-free: new head outputs Cholesky Gaussians for rotation","Angle continuity from Cholesky: GauCho beats DOTA baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000326,"raw_usage":{"total_tokens":1813,"prompt_tokens":918,"completion_tokens":895,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":820}},"tokens_in":534,"tokens_out":895,"duration_ms":8776,"temperature":1.0,"reasoning_tokens":820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:57:00.953653+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train an otherwise identical detector to regress the Cholesky parameters (α, β, γ) directly from a Gaussian ground truth, using a Gaussian loss, and plot the loss as a function of predicted orientation for a fixed anisotropic object. If a second local minimum persists near the boundary (e.g., near -90 degrees for a target at 89 degrees) that is almost as deep as the global minimum, the causal claim that the mapping is the culprit would be refuted; if the landscape becomes single-minimum, the claim survives.","supporting_citations":[{"cited_title":"Probabilistic intersection-over- union for training and evaluation of oriented object detectors","cited_arxiv_id":null,"evidence_quote":"Defines the ProbIoU loss and the OBB-to-Gaussian mapping, and introduces the encoding ambiguity for circular objects that GauCho addresses with oriented ellipses."},{"cited_title":"Fcos: Fully convolutional one-stage object detection","cited_arxiv_id":null,"evidence_quote":"The FCOS anchor-free detector that the GauCho head is built on and that shows the biggest DOTA gains."},{"cited_title":"DOTA: A large-scale dataset for object detection in aerial images","cited_arxiv_id":null,"evidence_quote":"The DOTA dataset used for the main experimental comparisons."},{"cited_title":"Rethinking boundary discon- tinuity problem for oriented object detection","cited_arxiv_id":null,"evidence_quote":"Recent work that notes Gaussian-based losses can still suffer boundary discontinuity, providing the motivation for the paper's causal hypothesis."},{"cited_title":"Rethinking rotated object detection with gaussian wasserstein distance loss","cited_arxiv_id":null,"evidence_quote":"The GWD loss, one of the Gaussian-based regression losses tested with GauCho."},{"cited_title":"Learning high-precision bounding box for rotated object detection via kullback- leibler divergence","cited_arxiv_id":null,"evidence_quote":"The KLD loss, another tested Gaussian-based loss, and the loss used in the paper's loss-landscape illustration."},{"cited_title":"Detecting rotated objects as gaussian distributions and its 3-D generalization","cited_arxiv_id":null,"evidence_quote":"The prior work that maps OBBs to Gaussian distributions and provides the anchor-based regression scheme that GauCho adapts."},{"cited_title":"On boundary discontinuity in angle regression based arbitrary oriented object detection","cited_arxiv_id":null,"evidence_quote":"Recent work on boundary discontinuity in angle regression that, together with [27], motivates the paper's claim about the residual discontinuity."},{"cited_title":"Learning roi transformer for oriented object detection in aerial images","cited_arxiv_id":null,"evidence_quote":"The RoI-Transformer two-stage detector that, paired with GauCho, yields the best DOTA result."}],"review_version":1}