acceptodds
Withdrawn from ICLR 2027

Reliable Object Detection Through Guided Post-Hoc Uncertainty Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable uncertainty estimation is essential for deploying object detectors when distribution/covariate shift and adversarial attacks may occur. Existing approaches often require detector retraining, architectural modification, or repeated inference, which may be infeasible or incur significant overhead. We introduce a lightweight post-hoc evidential meta-model that learns when object localisations should be considered uncertain while keeping the base detector frozen. Our approach automatically identifies localisation-relevant features and uses saliency-guided modification to construct an increasingly challenging curriculum. Detection-level targets combine localisation error, modification level, and prediction instability to guide an evidential meta-model to estimate uncertainty for each predicted bounding box. Our approach requires no changes to the detector and preserves its original localisation outputs. Across adversarial attacks and evaluated strengths, GRACE improves TP-FP AUROC by 22% relative to the strongest comparator in some cases while maintaining in-distribution detection performance.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Object Detection, Uncertainty Quantification, Evidential Regression, Post-hoc Uncertainty Estimation, Adversarial Robustness, Covariate Shift

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

The paper introduces GRACE (Guided evidential Regression for Adversarial and Covariate uncertainty Estimation), a localisation-aware post-hoc evidential meta-model for object detection that provides uncertainty estimates for each predicted object. One of the benefits of GRACE is the absence of retraining the model for object detection. The current approach was evaluated based on a single object detector (Yolov11n) against 4 datasets and 3 baselines model.

Critical strengths

I believe GRACE meta-model approach has its merit: having a meta-model that works for all object detection model is a valuable contribution. GRACE shows some promising results on CelebA and VisDrone dataset. However, the current approach is not generalizable to other object detectors and other datasets (performance-wise).

Critical weaknesses
  • The evaluation focus on a single object detector model (YOLOv11). Therefore, it is difficult to assess the generalizability of the proposed method.

  • Looking at the numbers from Table 1 and Table 3, GRACE does not always outperforms the baseline methods. In the case where GRACE outperforms the baseline methods, the performance gain is marginal. For Table 1, on the CelebA dataset, GRACE is vastly superior to baseline methods. However, things start to fall apart on the other datasets.

  • I do not think Figure 3, FIgure 4, and Figure 5 provide a truthful representation of GRACE performance (robustness) to (non-adversarial and adversarial) perturbations targeting the image. As seen in Table 1, under normal conditions (no perturbation), GRACE is performing well on CelebA compared to baseline methods. A better approach would be to show Figure 3, 4, and 5 on datasets such as TT100k or COCO where GRACE is on par with the baselines (in absence of perturbations).

Questions for authors
  • Have you evaluated GRACE using a model different than YOLOV11n?
  • How fast is GRACE on a edge device CPU (smartphone or laptop)?
Reviewer_Ly2zOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GRACE, which trains an evidential meta-model on top of a frozen YOLO11n detector. Layer selection and localization-relevant image regions are identified using LRP, and a saliency-weighted salt-and-pepper pixel perturbation strategy is used to construct a training curriculum. For each retained detection, the training objective combines localization error, the expected local perturbation ratio, and prediction stability before and after perturbation. During inference, the meta-model uses features extracted from bounding-box regions to predict NIG parameters without modifying the original detection boxes. The authors evaluate GRACE on COCO, CelebA, VisDrone, and TT100K, reporting results for detection uncertainty on clean data, common image corruptions, and adversarial attacks, including PGD, TOG, and Square Attack. The experiments use YOLO11n. The paper reports that GRACE outperforms EMM, ModelNet, and MetaDetect on several datasets and metrics, although the extent of its advantage varies across datasets and evaluation metrics.

Critical strengths
  1. The paper has a practical objective and reasonable deployment motivation. Estimating localization uncertainty on top of an existing detector while preserving its outputs offers practical value for applications in which retraining or modifying the underlying model is infeasible.
  2. Beyond distinguishing genuine from false detections, the paper also examines the relationship between uncertainty and localization quality, as well as the degradation of localization performance and the disappearance of detections under image perturbations.
Critical weaknesses
  1. The paper appears to extract ROI features from perturbed detection boxes, fix the NIG predictive mean to the clean-image boxes, and supervise regression with matched ground-truth boxes. At the same time, localization error and stability targets are constructed by matching clean and perturbed detections. It is difficult to tell from the current description how these quantities fit together. Please specify which boxes define the ROIAlign regions, which detections are matched to ground truth, and which boxes serve as the predictive means. Please also state whether the NIG regression loss is computed when a clean detection has no match in a perturbed view. If the interpretation above is correct, the paper should explain why uncertainty predicted from perturbed-view features reflects the relevant localization error. If it is incorrect, the equations and algorithm should be revised to match the implementation. This clarification is essential to assessing the uncertainty targets.
  2. The reported ranking and localization correlation results show that the uncertainty scores may be useful, but they do not separately validate the proposed aleatoric and epistemic interpretations of the NIG parameters. The paper should either test these interpretations directly or limit its claims accordingly. The comparative results also warrant a more careful discussion. EMM performs better on several TT100K calibration metrics, so GRACE’s advantage is not consistent across the reported measures. The paper should explain how the geometry-only MetaDetect variant was constructed, how it differs from the original method, and why it is a suitable baseline. Claims of superiority should reflect the full set of results. All main experiments use YOLO11n. Evaluating at least one detector with a different architecture would strengthen the claim that GRACE transfers across detectors; otherwise, that claim should be narrowed. Finally, the cost analysis should include the extra frozen-detector passes used during training and the LRP computations used for saliency calibration, alongside measured inference overhead. Separating one-time calibration costs from costs incurred during deployment would give readers a clearer picture of the method’s practical value.
Questions for authors
  1. A worked example would make the pipeline much easier to follow. Please identify the clean detection box, its matched box in the perturbed view, the box used for ROIAlign, the ground-truth box, and the fixed NIG mean. The paper should also explain what happens when a detection disappears under perturbation, and whether Eqs. 12, 14, and 16 and Algorithm 1 use the same detection indices. Without this information, it is hard to judge whether the loss terms supervise the same prediction or whether the resulting uncertainty reflects localization error.

  2. During training, qgtq^{gt} uses ground-truth annotations, while ucorru^{corr} uses the known perturbation probability pt(h,w)p_t(h,w). Neither is available in the same form at inference. Please state which inputs the uncertainty head receives at test time and how, if at all, these training signals are replaced. Evaluation on natural distribution shifts or perturbations beyond salt-and-pepper noise would also help establish what the method has learned. As presented, the results may show sensitivity to the training perturbation and its effect on detection quality, but do not yet establish general OOD detection or a reliable estimate of epistemic uncertainty.

  3. The appendix studies sensitivity to TT, γ\gamma, and η\eta, but does not isolate the contributions of saliency-guided perturbation, curriculum learning, or evidential learning. Useful comparisons would include uniform instead of saliency-guided perturbations, training without the curriculum, separate use of the localization-error and perturbation-exposure/stability targets, and a deterministic head in place of the evidential head. These ablations would show which components account for the reported gains.

  4. Eq. 1 uses the sum of four raw box coordinates as the LRP attribution target. Please specify the coordinate format, normalization, and treatment of output biases. It would also be useful to compare corner coordinates with center–width–height coordinates and examine whether coordinate scale or a different detection head changes the layer rankings or saliency maps. This would show how broadly the saliency-based feature selection can be applied.

  5. Please explain why MetaDetect is evaluated as a geometry-only variant and how removing its other features may affect the comparison. The detector, training data, detection thresholds, matching rules, and hyperparameter selection used for each method should also be stated. For the efficiency claim, report calibration cost, meta-model training cost, any extra detector or attribution passes, and end-to-end inference latency separately. In particular, clarify whether detector inference or attribution is repeated for each training sample, and quantify that cost.

Reviewer_SeVXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Guided Evidential Regression for Adversarial and Covariate Uncertainty Estimation (GRACE), which is a post-hoc uncertainty estimator for object detection. It learns when object localizations should be considered uncertain while keeping the base detector frozen. The framework is reasonably designed, with substantial improvements on CelebA and promising results on VisDrone.

Critical strengths
  • Estimating localization reliability without retraining the underlying detector is a useful and well-motivated objective.
  • The evaluation covers four datasets and includes TP-FP AUROC and FP AUPRC for incorrect detection identification, risk-coverage and IoU-coverage for selective prediction, and the Spearman correlation between uncertainty and detection IoU.
  • The CelebA results show substantial improvements, and the VisDrone results provide additional evidence that the method can learn useful localization-reliability information.
Critical weaknesses
  • YOLO11n is used in the experiments. However, this is not sufficient to establish generality across different detector architectures. At least one substantially different detector family would help determine whether this approach can generalize.

  • The work by Barker et al. (cited in the paper) already combines frozen base model, extraction of salient layers, and curriculum generation. This paper introduces detection-specific adaptations. However, the paper should more clearly establish which adaptations provide new value.

  • The motivation involves closed-weight models. Yet, the proposed method requires intermediate activations and relevance propagation. This limits the deployment settings that the current method supports.

  • The proposed approach should be evaluated on the SAOD benchmark (NeurIPS 2023, “Towards Building Self-Aware Object Detectors”).

  • The current comparisons and ablations do not sufficiently establish the method’s novelty, the necessity of its components, or the generality of its benefits.

Questions for authors

Please see the items listed under Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.