acceptodds
Withdrawn from ICLR 2027

EAGLE: Entropy-Aware Geometric Learning Enhancer for Partially Polygon-Supervised Object Localization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Object detection is a critical application of computer vision, yet models trained uniformly on bounding boxes lack the geometric precision to be fully effective in inference settings, especially at high IoU thresholds. One common approach is the introduction of polygon-supervised localization, where a model is trained explicitly for localization but with access to polygon masks during training. Among such methods, partial polygon-supervision is viable due to the higher annotation costs of developing fully supervised training datasets. However, this introduces a further challenge: in object classes of relatively low shape complexity, the addition of training masks and subsequent segmentation loss may obscure box localization, often worsening model performance despite access to deeper geometric information. To address the practical limitations of both full and partial polygon supervision, we introduce a mathematical framework for quantifying per-class geometric complexity by combining information theory and statistical shape analysis. With this framework, we develop Entropy-Aware Geometric Learning Enhancer (EAGLE), a lightweight, post-training model head that learns geometric information for the localization model, improving accuracy via edge refinement and localization rescoring using entropy-informed updates. Empirical analysis across MS COCO and Pascal SBD demonstrates the entropy framework meaningfully predicts shape reconstruction error and accuracy differences among mask-derived and standard box predictions, while the implemented EAGLE framework achieves consistent improvement on already SOTA detection frameworks (+0.9 mAP50-95 points on MS COCO) at minimal observed inference overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
object detection, semi-supervised learning, computer vision, instance segmentation, information theory, statistical shape analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_mA2iOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes EAGLE, which uses shape entropy computed from polygon-annotated samples to quantify an object's intrinsic geometric complexity, and uses this measure to refine bounding-box predictions. The claimed contribution is improved box localization at comparable latency. Validation consists of comparisons against YOLO26 baselines (e.g., n and m variants), reporting AP and latency; the full EAGLE pipeline reports 41.11 AP (n) and 52.94 AP (m).

Critical strengths
  1. Reasonable motivation. Explicitly quantifying object shape complexity and using it as a prior for bounding-box regression is an intuitive idea, and shape entropy is a sensible, lightweight descriptor for this purpose. This is the main point in the paper's favor. However, it carries limited weight in my assessment, because the paper does not provide convincing evidence that the idea actually works (see Weaknesses).
Critical weaknesses
  1. The empirical evidence does not support the claimed improvement. The YOLO26 baselines are incompletely reported. With the official default NMS head and no additional training, YOLO26n achieves 40.9 AP, only ~0.2 below the full EAGLE pipeline (41.11), and YOLO26m achieves 53.1 AP, which exceeds EAGLE (52.94). The remaining differences in AP and latency are small enough to fall within run-to-run noise, and no variance or multi-seed results are reported. Furthermore, there is no ablation isolating the contribution of the shape-entropy conditioning, nor any study of the proportion of polygon-annotated samples. This is consequential because the paper's central claim is an empirical improvement over the baseline; against properly configured baselines, that improvement is marginal or negative, and there is no evidence that the proposed component is responsible for any gain.

  2. The technical content and motivation lack rigor. Several equations are incorrect or unjustified: e.g., in C(r) (Line 211) the denominator contains rir_i with a dummy index ii but the summation ∑i\sum_i is missing, and Line 307 defines Hi=0.8⋅(…)+0.2⋅(…)H_i = 0.8 \cdot (\ldots) + 0.2 \cdot (\ldots) with unexplained constants and no sensitivity analysis. More seriously, the paper's motivating premise misrepresents prior work: Related Work cites Mask R-CNN as an example of "geometric feature interference," claiming that joint mask training harms box localization, whereas the original paper (He et al., 2017, Table 3) reports that multitask training improves box AP by +0.9. Presentation is also well below the venue standard (a single figure with illegible text at Line 386; a table whose Latency column overflows the page at Line 441). This is consequential because, beyond weak results, the paper's theoretical formulation and its stated motivation cannot be trusted as written.

Questions for authors

All questions below are clarification requests whose answers could change my assessment; I have no suggestions for future work at this stage.

  1. Baselines. Why do the reported YOLO26 baselines differ from the official results (40.9 AP for YOLO26n and 53.1 AP for YOLO26m with the default NMS head)? Please report baselines under the official configuration, along with mean ± std over multiple seeds for both EAGLE and the baselines, and specify the latency measurement protocol (hardware, batch size, precision, warm-up, number of runs). If EAGLE shows a statistically significant gain over correctly configured baselines, I would reconsider my assessment.

  2. Contribution of shape entropy. What is the performance of the full pipeline with the shape-entropy term removed (or replaced by a constant), and how does performance vary with the proportion of polygon-annotated training samples? Without these results, it is unclear whether the proposed component contributes anything.

  3. Mask R-CNN claim. What evidence supports the claim that joint mask training harms box localization, given that the original Mask R-CNN paper reports the opposite (+0.9 AP^bb from multitask training)? If this premise does not hold, how is the method still motivated?

  4. Equations and constants. Please correct the C(r) definition (Line 211) and explain how the 0.8/0.2 weights (Line 307) were chosen, including a sensitivity analysis.

  5. Authorship process and verification. Taken together, the incorrect equations, unexplained constants, unpolished figures and tables, and the mischaracterization of Mask R-CNN suggest that substantial portions of the paper may have been produced with an automated research pipeline or LLM-based tools without adequate human review. Could the authors clarify which parts of the work, if any, were produced with such tools, and what steps were taken to manually verify the equations, experimental results, and citations?

Reviewer_ACugOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes to quantify the shape complexity for the objects within every object class, obtaining the shape entropy. Using this shape entropy, EAGLE is proposed to refine the box localization. The paper mainly introduces theoretical formulation of how this shape entropy derives.

Critical strengths

Quantifying the shape complexity may be useful to enhance the object localization. The paper gives a potentially valid way by calculating the shape entropy for the object classes.

Critical weaknesses
  1. The experiment is insufficient and the effectiveness of the proposed method is not convincing. The method EAGLE utilizes shape entropy to guide the localization learning and inference. However, the figures in Appendix D.1 show that the fitted lines are positive in three of the six subplots, indicating that mask supervision does no harm to those classes having a low shape entropy. When the shape entropy is large, there are still many points below 0. In addition, the improvements are relatively marginal. Table 1 shows that the proposed method only boosts AP by 0.45 over YOLO26m and 0.16 over RT-DETRv1-S. Another concern is about the generalization ability. The evaluations are conducted on two YOLO26 models, one RT-DETRv2 model, and COCO 2017val, which is insufficient. The proposed method requires a certain amount of polygon annotation data to work. This is expensive to get and the prior knowledge of COCO can probably not be generalized to other domains, e.g., remote-sensing, SAR, and drone scenarios. More detectors and datasets are required to demonstrate the generalization ability.Further more, the paper mainly study the close-set detection. It would be invalid if class is open-vocabulary. Since the method is not a dynamic feature learned by the model, but an external statistical prior, which indicates that the method design is data-specific and less generalizable.

    The paper also claims that a det+seg framework is likely inferior to the pure det framework because the gradient of the mask branch can interfere with detection. However, this argument overlooks two important aspects. First, a joint det+seg model produces instance masks as an additional output, which is inherently unavailable from a detection-only model. Second, substantial prior literature has demonstrated that properly designed multi-task detection-segmentation frameworks can improve box localization and detection accuracy rather than degrading it. The negative transfer observed in YOLO26-seg may be a property of that specific YOLO26 multi-task implementation and training recipe, rather than a general conclusion for all det+seg systems. Without a thorough discussion of this multi-task learning, the paper's general statement about the disadvantages of the det+seg paradigm is one-sided and misleading.

  2. The method involves complicate designs on the computation flow. Multiple hyper-parameters are set without ablation studies. For example, why use 40% of training polygon instances? The method may need to carefully tune the hyper-parameters, which undermines the robustness of the method.

Questions for authors

More evaluation and ablation studies are needed.

Reviewer_YdUAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces EAGLE, an entropy-aware geometric learning enhancer. With only 40% polygon supervision, the authors quantify category-level conditional shape entropy H(S∣Abox)H(S|A_\mathrm{box}) via Kendall shape space and information theory, then train a lightweight post-hoc head to refine edges, rescore predictions and post-process outputs from a frozen detector. Empirically, it yields consistent but small mAP50-95 gains with negligible inference cost (e.g., +0.9 on YOLO26n). The main theoretical contribution is a framework for category-level shape uncertainty under partial polygon annotations, formalizing H(S∣Abox)H(S|A_\mathrm{box}) and deriving entropy bounds alongside an expected IoU bound that decays with shape uncertainty. Correlation analyses are provided to argue that shape entropy predicts reconstruction error and detection performance uplift.

My major concerns are threefold. First, the comparisons against prior art are insufficient, making it hard to contextualize the proposed method. Second, the empirical improvements are marginal, which weakens the practical value of the proposed entropy machinery. Third, the manuscript suffers from poor polish and inconsistent formatting, suggesting insufficient revision before submission.

Critical strengths
  1. The work identifies a nontrivial observation: polygon supervision at the category level with limited complexity can interfere with box localization. This phenomenon is evidenced by feature interference in joint training of Mask R-CNN and the degraded box AP of YOLO with instance segmentation compared to its detection-only counterpart. Rather than naively stacking an additional segmentation head, the authors frame this observation as a quantifiable and solvable problem.
  2. It pioneers the integration of information theory and statistical shape analysis (Kendall’s shape space) for annotation-level embedding (operating at the box/polygon level instead of the image level), deriving computable category-wise shape entropy. This fills a genuine gap in the literature. Experiments and comparisons are conducted across both CNN-based and transformer-based architectures.
Critical weaknesses
  1. Empirical results show that the proposed method yields only modest gains over the baselines. The comparison set is limited: the authors do not compare against recent state-of-the-art approaches, nor do they include strong baselines equipped with common augmentation strategies such as multi-scale training. Such augmentations often narrow performance gaps and may invalidate the observed improvements.
  2. The core validation for the claim that “entropy predicts localization advantage” is weak. Section 7.2 provides the key causal evidence for whether entropy can guide localization, yet the average correlation coefficient is only 0.298. It reaches 0.540 only after outlier removal. The most critical supporting result relies heavily on outlier filtering, which undermines its persuasiveness. The paper mainly shows that higher shape uncertainty imposes a lower upper bound on box accuracy. This is a useful conceptual motivation but offers no actionable design rule, nor a guarantee that reducing entropy improves detection performance. It is unclear whether the +0.9 mAP50-95 gain comes from the entropy signal or merely the added lightweight head and post-processing, as no ablation disentangles these effects.
Questions for authors
  1. Table 1 exceeds the page width. Besides, the text inside Figures 1, 2 and other figures is too small to read clearly. Please re-fit Table 1 to the page width and enlarge the text in Figures 1 and 2.
  2. Why is the comparison scope limited, without SOTA methods and strong baselines with multi-scale augmentation? Improvements from auxiliary modules often disappear under stronger augmentation.
  3. The core correlation in Sec. 7.2 is weak (ρ\rho=0.298) raw, 0.540 after outlier removal). Please justify the outlier filtering and provide more robust evidence for the entropy-localization claim.
  4. Can you separate the mAP gain from the entropy signal versus the extra head and post-processing? Is there an ablation without entropy conditioning?
  5. The entropy bound only establishes an upper-limit constraint rather than a monotonic guarantee: lowering uncertainty may not improve AP. Could you provide more evidence for this causal link?

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

Related papers

Discussion (0)

Sign in to comment.