{"id":"c2ce42c6-3dbb-4566-8305-57b195911341","arxiv_id":"2411.10497","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Encoding oriented bounding boxes as structure tensors and regressing the three matrix entries with L1 loss gives accurate, low-cost angle prediction in oriented object detection.","lead":"This paper encodes an object's orientation as a 2x2 structure tensor and regresses it with L1 loss, avoiding angle-periodicity problems. On five detection benchmarks, this simple representation matches or beats prior angle encodings in mean average precision and angular error.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central construction is internally inconsistent for square-like OBBs: Section 4 changes the encoder to lambda1=w, lambda2=h/2 for |w-h|<=epsilon, while Eq. 5 still decodes w=2*lambda1, so a square of side s becomes 2s x s.","rationale":"The reader's weakest_assumption identifies exactly this square-object encode/decode inconsistency, and I agree it is the most load-bearing issue. The novelty claim for structure tensors depends on the representation being invertible; if the decoder is wrong for the very symmetric objects the authors emphasize, the mAP gains cannot be credited to the described method. The concern is not merely a missing hyperparameter (epsilon) but a missing mathematical case in the decoder. It is independently testable by inspecting the claimed public code, so the existing CONDITIONAL verdict is the right level of confidence. My read does not change that verdict: if the round-trip check confirms the mismatch, the paper would need rejection or major revision, but until that check is run the appropriate posture is conditional acceptance pending clarification.","tokens_in":13960,"tokens_out":8165,"duration_ms":85129,"concrete_test":"Extract the released encoder and decoder (or reimplement Eqs. 3-5 exactly) and round-trip every ground-truth OBB in DOTA and C19TD that satisfies |w-h|<=epsilon, recording decoded w, h, and theta versus GT. For a square of side s, the published decoder should output w=s, h=s if a hidden rule exists; if it outputs w=2s, h=s, the reported tables were produced by an undocumented decoder and the central reproducibility claim fails. Use the epsilon value from the code if present; otherwise test epsilon=0 and epsilon=1 to bracket the ambiguity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim requires the encoding and decoding steps (Eqs. 3-5) to be mutually inverse for every OBB in the evaluation set. Section 4, 'Handling Isotropic Objects,' breaks this for square-like boxes. For a square with side s, the default encoder gives lambda1=lambda2=s/2, yielding no principal direction. To create a 2:1 anisotropic target, the text sets lambda1=w=s and lambda2=h/2=s/2, and normalizes the ground-truth angle into a pi/2 interval when |w-h|<=epsilon. The decoder in Eq. 5 remains w=2*lambda1, h=2*lambda2, so an exact prediction of T_gt decodes to w=2s, h=s rather than the original square. Thus either an undocumented second decoding rule is used in the implementation, or square-like objects are systematically misrecovered at inference. The threshold epsilon is never specified, and no code link is provided despite the abstract's claim, so readers cannot check which branch occurred. Because DOTA (storage-tank and roundabout classes) and C19TD are explicitly chosen for their symmetric, square-like objects, the reported numbers depend on resolving this inconsistency. This is not a tuning issue; it is a mismatch between the published mathematical description and the evaluation pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to represent the orientation of an oriented bounding box (OBB) by a 2x2 structure tensor T = R_theta * diag(lambda1, lambda2) * R_theta^T, where the network regresses the three parameters (a, b, c) under an L1 loss. The encoding/decoding pair is given by Eqs. (3)-(5), mapping (w, h, theta) to T and back. To handle square-like objects, Section 4 modifies the encoder for |w-h| <= epsilon to set lambda1 = w and lambda2 = h/2, and normalizes the angle to a pi/2 range. The method is evaluated on DOTA, HRSC2016, ICDAR2015, MSRA-TD500, and a new C19TD dataset, with comparisons to CSL, GWD, KLD, KFIoU, PSC, and others, plus ablations on the anisotropy ratio and computational cost.","tokens_in":14264,"tokens_out":3129,"duration_ms":28688,"significance":"The structure tensor representation is a simple, modular idea that plausibly combines the benefits of angle-coder methods (low computational overhead, direct regression) with the robustness of Gaussian-based representations to angular periodicity and symmetry. If the encoding/decoding inconsistency for square-like objects is resolved, the approach could be a useful alternative in oriented object detection. The paper provides extensive empirical evaluation across five datasets, including a purpose-built dataset of symmetric COVID-19 test objects, and a direct comparison of angular precision (Table 3) that is a useful contribution. The computational cost analysis in Table 5 is also a strength. However, the central claim of consistently achieving state-of-the-art results is not supported by the paper's own Table 1, and the handling of isotropic objects is not self-consistent.","major_comments":[{"comment":"The handling of square-like objects is internally inconsistent. For |w-h| <= epsilon, the encoder is changed so that lambda1 = w and lambda2 = h/2, but the decoder in Eq. (5) still computes w = 2*lambda1 and h = 2*lambda2. For a square of side s, an exact prediction of the encoded tensor decodes to w = 2s, h = s, which is not the original box. No alternative decoding rule is described, and the threshold epsilon is never specified. Since DOTA's storage-tank and roundabout classes and the C19TD dataset are explicitly chosen for their square-like objects, the reported results depend on this step. Please clarify the actual decoding rule for these objects, specify epsilon, or revise the encoding so that Eqs. (3)-(5) are mutually inverse for all OBBs.","section":"Section 4, Eq. (5)"},{"comment":"The text claims that the approach 'achieves SOTA results for the DOTA dataset on OBB-supervised methods,' but Table 1 does not support this. On DOTA mAP50, KLD (71.6) and CSL (71.4) both outperform ST (71.3), and on mAP50:95, CSL (40.5) outperforms ST (39.5). The discussion should be revised to reflect the actual rankings rather than overclaiming state-of-the-art status.","section":"Section 5.2, Table 1"},{"comment":"The paper repeatedly states that the method requires 'no additional hyperparameters' or 'does not require hyperparameter tuning.' However, the anisotropic encoding for square-like objects introduces two free choices: the 2:1 anisotropy ratio and the threshold epsilon. The ablation in Table 4 selects the 2:1 ratio based on validation performance, which is a hyperparameter selection. This claim should be qualified or removed.","section":"Section 4 and Section 6"}],"minor_comments":[{"comment":"The abstract states 'We make our code publicly available,' but no code link is provided anywhere in the manuscript. Please include the repository URL.","section":"Abstract and Section 1"},{"comment":"The phrase 'To the best of out knowledge' contains a typo; it should be 'our knowledge.'","section":"Section 1"},{"comment":"The text says 'we handle the cases where w = h,' but the actual condition is |w - h| <= epsilon, not strict equality. Please rephrase for consistency.","section":"Section 4"},{"comment":"The formula for the angular error delta has a redundant nested min: min(min(|theta_pred - theta_gt|, |theta_pred - theta_gt + pi|), |theta_pred - theta_gt - pi|). Simplify to a single min over the three absolute values.","section":"Eq. (9)"},{"comment":"The header 'Angle MSE ( M AEθ)' appears to be a typo; it should be 'Angle MAE (MAEθ).'","section":"Table 3"},{"comment":"The C19TD dataset description mentions 'we highlight the significance of this data' but it is a new dataset; the introduction should make clear that this is a new contribution and describe how the dataset will be released.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The central inconsistency in Section 4 is load-bearing and must be fixed before publication; the authors need to either describe the actual decoding rule used in their implementation or adjust the encoding so that it is invertible. The SOTA overclaim in Section 5.2 is also a correctness issue in the presentation. The idea is potentially valuable and the experimental effort is substantial, so I recommend major revision rather than rejection. Please also ensure the code link is provided, as the abstract promises it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is clean: represent the orientation of an OBB as a 2x2 symmetric structure tensor and regress its three components with an L1 loss. This is a natural bridge between Gaussian-based methods and angle coders, and it ships as a modular angle head that costs almost nothing in FLOPs or parameters. The evaluation is extensive—five datasets, including a new C19TD benchmark, plus dedicated angular error metrics—and the results are broadly competitive, sometimes best. That part is genuinely useful; anyone building a rotated detector would want this baseline.\n\nThe serious soft spot is the square-object handling in Section 4. The encoder switches to lambda1 = w, lambda2 = h/2 when |w - h| <= epsilon, but the decoder in Eq. 5 still recovers w = 2*lambda1 and h = 2*lambda2. For a square of side s, an exact prediction would decode to 2s x s. No alternate decoding rule is given, epsilon is never specified, and the abstract promises public code but no link appears in the paper. Because DOTA and C19TD were explicitly chosen for their symmetric objects, this ambiguity sits right on top of the reported numbers. It might be a harmless implementation detail that the text just failed to document, but as written the method is internally inconsistent for square-like boxes.\n\nThe SOTA claim is also overstated. Table 1 shows KLD (71.6) and CSL (71.4) beating ST (71.3) on DOTA mAP50, and CSL beating ST on mAP50:95. That is not \"consistently achieving SOTA.\" The real story is that ST matches or slightly beats existing methods on most benchmarks, with lower cost and no extra hyperparameters. That is worth publishing, but the text should say it.\n\nMinor issues: no error bars, and the new C19TD dataset is mentioned but not released, which limits independent verification.\n\nWho is this for? Researchers working on oriented object detection who want a cheap, robust angle representation. It deserves peer review because the idea is sound and the flaw looks fixable rather than fatal—but the authors need to clarify the square-object decoder, specify epsilon, and ideally release the code and dataset. I would send it to review with a request for major revision on those points.","headline":"A simple, modular angle representation that mostly works, but the paper's own tables don't support the SOTA claim and the square-object handling has an undocumented encoding/decoding mismatch.","tokens_in":14814,"tokens_out":2183,"would_cite":false,"duration_ms":24582,"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":"The paper claims that encoding an oriented bounding box as a structure tensor—a 2×2 symmetric matrix with three parameters—and regressing those parameters with an L1 loss outperforms prior angle-coder and Gaussian-based methods across…","keywords":["oriented object detection","structure tensor","angle representation","angular boundary problem","symmetry ambiguity","bounding box regression","weakly supervised detection","remote sensing"],"falsifier":"Run a reconstruction test: encode a square box with the anisotropic rule, apply the paper's decoder, and compare the output (w,h,θ) with the input. If the decoded width is 2w, then the representation as specified cannot faithfully reconstruct square boxes, and the reported mAP improvements on datasets containing squares would need to be re-examined for whether the detector is ignoring the width error.","tokens_in":13748,"feed_emoji":"🧭","tokens_out":8077,"duration_ms":75746,"temperature":0.7,"pith_summary":"Oriented object detection adds a rotation angle to the usual bounding box, but predicting that angle directly is brittle: near the angular boundary tiny errors cause large losses, and symmetric objects make different angles produce the same box. This paper proposes to stop predicting the angle at all, and instead predict the three parameters (a, b, c) of a 2×2 structure tensor that encodes the box's width, height, and orientation through its eigenvalues and eigenvectors. Training with a plain L1 loss on those parameters makes equivalent rotations map to nearby tensors, so the boundary and symmetry problems largely disappear without extra hyperparameters. Across five datasets—aerial imagery, scene text, and a new COVID-19 test-strip benchmark—the representation reaches state-of-the-art or better mean average precision, with angular errors among the lowest reported and computational overhead near the bottom of the comparison.","feed_headline":"Structure tensor beats angle coders on five rotated-object datasets","feed_subtitle":"Encoding box angle as a 2x2 symmetric matrix removes boundary and symmetry training losses at near-zero extra cost.","key_machinery":"The load-bearing object is the structure tensor T = Rθ Λ Rθ^T with Λ = diag(w/2, h/2), originally used in edge and corner detection; its three independent parameters (a, b, c) are the network's regression targets. The eigenvectors of T carry the box orientation, the eigenvalues carry width and height, and the continuity of T as a function of θ removes angular boundary cliffs; the eigenvalue ratio handles the symmetry of rectangles (π-periodic) and squares (π/2-periodic). Encoding and decoding are simple algebraic operations—no distributional distance, no hyperparameters—so the representation drops into an existing detector's angle head by replacing the scalar angle target with a three-vector target and applying an L1 loss.","core_discovery":"In the proposed representation, a ground-truth oriented box (x, y, w, h, θ) is encoded as T = Rθ diag(w/2, h/2) Rθ^T, a symmetric matrix with components a, c, c, b. The angle head of a one-stage detector regresses these three numbers with an L1 loss; at inference the angle is read back from the eigenvector belonging to the larger eigenvalue, and width and height from twice the eigenvalues. Because the tensor entries vary continuously with rotation, the loss landscape no longer contains the sharp cliffs at ±π/2 or around the angular boundary that plague direct angle regression, and the eigenvalue ratio encodes the object's symmetry class so that rectangular and square objects are penalized consistently. Comparisons on DOTA v1.0, HRSC2016, ICDAR2015, MSRA-TD500, and the new C19TD dataset show that this representation matches or beats prior angle-coder and Gaussian-based baselines on mAP, and dedicated angular-error metrics (MAEθ and RMSEθ) place it among the most precise predictors. For square-like objects the encoder deliberately breaks isotropy by setting eigenvalues to w and h/2 (a 2:1 anisotropy) and normalizing the angle to a π/2 range, which the authors show improves both angular accuracy and mAP.","pith_inferences":["Because the tensor loss treats a box and its π-rotated equivalent as identical, the representation could transfer to other continuous periodic targets, such as oriented keypoints or vanishing points, wherever direct regression suffers boundary cliffs.","The decoder described for the general case recovers width as 2λ1, while the anisotropic square encoding sets λ1 = w; reconciling these two rules would require either a decoding threshold or a small width correction, a detail the paper does not specify.","A natural extension is to let the anisotropy ratio depend on the object's aspect ratio rather than a fixed 2:1 threshold; the paper's ablation suggests a continuum worth exploring.","The angular-error metric used in the paper's evaluation (minimum over π-shifted differences) treats rectangles as undirected segments; adopting a strictly oriented metric would test whether the tensor's symmetry handling aids oriented accuracy beyond the un-oriented metric."],"forward_implications":["Replacing scalar angle regression with three-parameter tensor regression removes the need for hand-crafted angle losses or extra hyperparameters in rotated detectors.","Fully supervised (OBB) detectors gain mAP on aerial and scene-text benchmarks; weakly supervised (HBB-only) detectors also improve, suggesting the tensor gives the network a learnable orientation signal even without angular labels.","Angular precision, measured by MAE and RMSE of the predicted angle on true positives, is competitive with or better than the best angle coder tested, at similar computational cost.","The 2:1 anisotropic encoding for near-square objects improves angle error and mAP over isotropic encoding, and a stronger 4:1 anisotropy further lowers angle error but reduces mAP, so the anisotropy level is a tunable property of the representation."],"supporting_citations":[{"why":"This reference introduced the structure-tensor concept that the proposed angle representation is built on.","marker":"[1]"},{"why":"This work defined the circular smooth label angle-coder baseline that the paper compares against on boundary problems.","marker":"[40]"},{"why":"This work introduced the Gaussian Wasserstein distance, a Gaussian-based rotation representation used as a primary baseline.","marker":"[45]"},{"why":"This work proposed the Kullback-Leibler divergence for rotated boxes, another Gaussian-based baseline the paper compares against.","marker":"[46]"},{"why":"This work introduced KFIoU, a parameter-free Gaussian-based loss that is among the strongest Gaussian baselines tested.","marker":"[49]"},{"why":"This work proposed the phase-shifting coder, the closest predecessor that regresses angles through a transformed representation with an L1-style loss.","marker":"[53]"},{"why":"This work provided the H2RBox-v2 weak-supervision framework in which the structure-tensor encoder is evaluated on HBB-supervised tasks.","marker":"[55]"}],"fun_headline_variants":["Structure tensor angle encoding beats angle-coder baselines on five datasets","Structure tensor for oriented boxes: no angle cliffs, no extra hyperparameters","Angle regression via structure tensor: robust to periodicity on five datasets","Structure tensor angle head: simple, symmetric, and high precision","Encoding box angle as a tensor matrix solves boundary and symmetry issues"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For square-like objects (|w−h| ≤ ε) the encoder breaks isotropy by using eigenvalues λ1=w and λ2=h/2, but the paper's decoder formula w=2λ1 would then return twice the true width, and no separate decoding rule or value for ε is given, so the method assumes this mismatch does not hurt the reported results.","fun_headline_variants_meta":{"raw":{"variants":["Structure tensor angle encoding beats angle-coder baselines on five datasets","Structure tensor for oriented boxes: no angle cliffs, no extra hyperparameters","Angle regression via structure tensor: robust to periodicity on five datasets","Structure tensor angle head: simple, symmetric, and high precision","Encoding box angle as a tensor matrix solves boundary and symmetry issues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000933,"raw_usage":{"total_tokens":4001,"prompt_tokens":961,"completion_tokens":3040,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":2950}},"tokens_in":577,"tokens_out":3040,"duration_ms":24961,"temperature":1.0,"reasoning_tokens":2950,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:00:21.599450+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a reconstruction test: encode a square box with the anisotropic rule, apply the paper's decoder, and compare the output (w,h,θ) with the input. If the decoded width is 2w, then the representation as specified cannot faithfully reconstruct square boxes, and the reported mAP improvements on datasets containing squares would need to be re-examined for whether the detector is ignoring the width error.","supporting_citations":[{"cited_title":"Bigun, G.H","cited_arxiv_id":null,"evidence_quote":"This reference introduced the structure-tensor concept that the proposed angle representation is built on."},{"cited_title":"Arbitrary-oriented object de- tection with circular smooth label","cited_arxiv_id":null,"evidence_quote":"This work defined the circular smooth label angle-coder baseline that the paper compares against on boundary problems."},{"cited_title":"Rethinking rotated object detection with gaussian wasserstein distance loss","cited_arxiv_id":null,"evidence_quote":"This work introduced the Gaussian Wasserstein distance, a Gaussian-based rotation representation used as a primary baseline."},{"cited_title":"Learning high-precision bounding box for rotated object detection via kullback- leibler divergence","cited_arxiv_id":null,"evidence_quote":"This work proposed the Kullback-Leibler divergence for rotated boxes, another Gaussian-based baseline the paper compares against."},{"cited_title":"The kfiou loss for rotated object detection","cited_arxiv_id":null,"evidence_quote":"This work introduced KFIoU, a parameter-free Gaussian-based loss that is among the strongest Gaussian baselines tested."},{"cited_title":"Phase-shifting coder: Predicting ac- curate orientation in oriented object detection","cited_arxiv_id":null,"evidence_quote":"This work proposed the phase-shifting coder, the closest predecessor that regresses angles through a transformed representation with an L1-style loss."},{"cited_title":"H2rbox-v2: Incorporating symmetry for boost- ing horizontal box supervised oriented object detection","cited_arxiv_id":null,"evidence_quote":"This work provided the H2RBox-v2 weak-supervision framework in which the structure-tensor encoder is evaluated on HBB-supervised tasks."}],"review_version":1}