acceptodds
Under review as a conference paper at ICLR 2027

Rotation Sensitivity in Oriented Detectors Is Not One Quantity

Abstract

Oriented object detectors can be accurate without being equivariant to rotations of their input, and this lack of equivariance is usually summarised by a single number. We split the AP cost of the four axis-aligned views into a quarter-turn mode C, a half-turn mode H and a signed chirality Q, three fixed combinations of the discrete Fourier coefficients of per-view AP, and we measure the disagreement between predictions on different views separately from its AP cost. Both measurements need an exact map back to the original frame. Mapping predictions back with the pixel-index form of the inverse rotation is off by one pixel for continuous box coordinates, displaces the half-turn view along both axes and inflates H, so we check the mapping against the array rotation it inverts. In a 2×2 augmentation factorial on DOTA with paired seeds, each mode is removed by the augmentation that covers the rotations it measures. Adding the vertical flip, whose composition with the horizontal flip is the half turn, removes H (macro AP50 0.0218 → 0.0005) but leaves most of C (0.0505 → 0.0418), whereas adding rotation removes both (C = −0.0049, H = −0.0015). Bootstrap intervals over seeds and scenes separate H from zero only in cells trained without the half turn, and the pattern holds at a second model capacity, a second input size and a stricter overlap threshold. A control with rotations of at most 15° improves accuracy but removes neither mode. After rotation training no rotated view has a detectable AP cost, although, averaged over classes, more than a tenth of the predictions have no counterpart when the canonical and a rotated view are matched in either direction. A detector with a rotation-equivariant backbone and the same flips keeps about a third of the C that flips alone leave in our detectors (0.0132 against 0.0418). Because a half turn leaves box orientations unchanged, an explanation based on box orientation can only account for C. On DOTA, classes that are more anisotropic under a quarter turn lose more C under rotation training (Spearman ρ = −0.62, permutation p = 0.018), and H shows no such relation. This class-level relation does not hold on SODA-A, where C is indistinguishable from zero. Some small residuals remain, including a half-turn cost in the equivariant detector and a negative chirality when rotation is added without the vertical flip, and we report the tests we ran on them.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
oriented object detection, rotation equivariance, rotation sensitivity, data augmentation, evaluation methodology, empirical study, aerial images, DOTA

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uoVLOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how much an oriented object detector’s accuracy changes when the input image is rotated by 90°, 180°, or 270° and the predictions are rotated back. Instead of reporting one average number, the authors split the accuracy loss into a quarter-turn cost C (average loss on the 90° and 270° views), a half-turn cost H (loss on the 180° view), and a chirality Q (difference between the two quarter turns), arguing that these respond to different parts of training. To test this, they train a YOLO11-OBB detector on DOTA, an aerial-image benchmark with 15 object classes, under four augmentation settings that differ only in whether vertical flips and rotation augmentation are used (horizontal flips are always on), with three seeds each. They find that adding vertical flips, which together with horizontal flips produce a 180° rotation, removes H but leaves most of C, while adding rotation augmentation removes both; in other words, each cost disappears only when training covers that rotation. They also show that even when rotation-trained models lose no accuracy on rotated views, about 12% of their predictions have no matching prediction on the rotated view, so accuracy under rotation and rotation-equivariance are different properties.

Critical strengths
  1. The experiment is designed so that the augmentation is the only thing that changes: All four training settings start from the same checkpoint, share every hyperparameter, and are compared within a seed, so a difference in C or H between two cells can only come from the augmentation.

  2. It catches a measurement trap that would mislead anyone using the usual metric. Two models can score identically on rotated images while producing different detections; the paper shows this is what happens after rotation training. So “accuracy doesn’t drop under rotation” cannot be interpreted as “the model is rotation-equivariant”. That is a nice clarification.

Critical weaknesses
  1. The abstract is very long with ~20 numbers and never states the one-line result or takeaway very clearly. The body reads like a lab notebook and is hard to follow.

  2. Training on a rotation removes the cost of that rotation, which the authors admit is what coverage predicts, and the advice (use vertical flips and rotation) is already standard. The non-obvious findings (negative chirality, ReDet’s half-turn cost, 12% disagreement after rotation training) are not explained properly, not sure if I am missing to understand these properly though.

  3. The vertical flip removes 17% of C under macro weighting but 58% under instance weighting, and the chirality reversal holds under macro only. The paper reports both weightings but never states which readers should treat as primary, and the abstract uses the macro numbers alone.

Questions for authors
  1. Please report C per class in the H and HV cells, and state which weighting you consider primary. If the flip-resistant part of C comes from two or three low-instance classes, then it becomes more about those classes than the general detector right?

  2. Could you run a cell with degrees ∈ {0, 90, 180, 270} exactly, with no interpolation? If coverage alone removes C it should match HR/HVR; if it does not, continuous rotation also contributes something beyond coverage

Reviewer_zKdnOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates rotation sensitivity in oriented object detectors and argues that a single aggregate metric is insufficient to characterize the effects of input rotations. The authors decompose the AP cost across four axis-aligned views into three quantities. They further distinguish AP-level robustness from prediction-level disagreement after mapping detections back to the original coordinate frame. The empirical study employs a factorial augmentation design on DOTA using YOLO-OBB models at two capacities, supplemented by experiments on SODA-A and a rotation-equivariant detector. The results suggest that vertical flipping substantially reduces the half-turn cost, whereas rotation augmentation is more effective at reducing the quarter-turn cost.

Critical strengths
  1. The paper makes an important distinction between maintaining detection accuracy under input rotations and preserving prediction-level equivariance.

  2. The 2×2 augmentation factorial design, paired random seeds, shared initialization, and evaluation at two model capacities provide a useful basis for comparing augmentation strategies.

  3. The authors explain how confusing discrete pixel-index coordinates with continuous box coordinates can introduce systematic localization errors, and validate their implementation with synthetic geometric tests.

  4. The paper examines unmatched predictions, localization overlap, orientation changes, and confidence changes, providing a more informative picture of rotation sensitivity than AP alone.

Critical weaknesses
  1. It remains unclear how the average quarter-turn AP cost relates to the corresponding fourier component after accounting for the half-turn contribution.
  2. It remains unclear whether the authors propose three rotation-sensitivity modes with independent explanatory power, or merely reparameterize the per-view AP losses across four views.
Questions for authors

Does the absence of a detectable AP loss suffice to prove that the rotation mode has been eliminated?

Reviewer_kYxKOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper highlights that a rotated-object detector's lack of rotation equivariance should not be reduced to a single metric. By testing four 90-degree rotations, the authors break down accuracy loss into three distinct components: a quarter-turn cost, a half-turn cost, and an asymmetry cost between quarter turns. Additionally, the (human?) authors correct a common off-by-one mapping error in existing benchmarks that artificially inflates the half-turn cost. Targeted Augmentations: Experiments on the DOTA aerial dataset seem to show that each cost component is resolved only by specific data augmentations—vertical flips eliminate the half-turn cost, whereas full rotation training eliminates both quarter- and half-turn costs. Equivariance vs. Accuracy: Even with optimal training, over 10% of predictions differ across rotated views, suggesting that equal accuracy across angles does not guarantee true rotation equivariance.

Critical strengths

None in my view: I feel this document was AI generated with a poor human prompt and no post-creation control. See criticisms below.

Critical weaknesses

This submission is clearly AI generated using a poor human promt. Therefore, I agree with the decision taken by the AC in the first places (see his/her comment of october 6th below). Proofs that this submission has to be desk-rejected are the following:

Writing is extremely poor:

  • Abstract reads as pure concatenation of empirical checks and trivial findings on indices not defined.
  • The introduction doesnt have a single reference to motivate the work. The reader is uncerting about the goal.
  • Pages 1–5 cite nothing: The first reference appears in Section 4, page 6. the abstract, the introduction and the method section mention DOTA and YOLO-OBB by name but give no references. Most citations sit on page 9, where the related-work section.
  • AP is never defined (is it Average Precision?)
  • The figure 1 comes before its definitions so a reader of the caption can’t decode the formulas yet

Given the poor writing I declare myself unable to give a conclusive assessment of this (in my view) poorly motivated and explained report.

Questions for authors
  • What kind of prompt did the authors used?.
  • What side information?
  • What AI generated the text?
  • Why no human controlled the result?

Comments on the paper

1 comment

Thank you for the revisions and clarifications. However, I still find the paper difficult to follow, and the exact right-angle training control is still incomplete, so I maintain my original score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.