acceptodds
Under review as a conference paper at ICLR 2027

Deciding Which Detections to Release: BlockCalibration for Open-Vocabulary Detection

Abstract

Open-vocabulary detectors let users locate objects through text, but a confident prediction for an absent category can become a false alert or an incorrect training label. Reliable use therefore requires deciding which detections to release while retaining useful objects. This decision is complicated by the detector itself: each query produces a dependent set of candidates whose size changes with the image, query, and score filtering. Pooling these candidates for calibration gives queries with more proposals greater weight and can fail to control the average error across queries. We propose block calibration, a postprocessing framework that aligns calibration with this query-level objective. The method keeps each image and its queries together, averages false-detection evidence within queries and across the image, and uses the resulting reference with the e-value Benjamini–Hochberg rule to select detections. Under exchangeability of complete images and their queries, it controls the expected false fraction averaged across queries, allowing score-dependent candidate counts and arbitrary dependence within an image. A maximum-reference variant also controls the expected false fraction in an image’s combined output. We further characterize the exact release condition, explaining how evidence strength, candidate count, and rank determine which objects survive. Experiments with GroundingDINO and YOLO-World on COCO, LVIS, and Open Images examine both high-precision release and its cost in object recovery. On COCO/GroundingDINO-tiny, the learned-evidence rule recovers 16.51% of objects at 98.12% pooled precision, more than doubling the recovery of binary evidence at the same nominal error level. Filtering and prompt-bank studies further show how candidate generation shapes the usefulness of error-controlled release.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-vocabulary detection, false discovery rate, e-values, multiple testing, block calibration, post-hoc calibration, selective prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_75RaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper applies the e-BH procedure to filtering predictions from frozen open-vocabulary detectors. It proposes block calibration to control the expected false discovery proportion averaged across queries. Extensive experiments with GroundingDINO and YOLO-World on COCO, LVIS, and Open Images compare the proposed method with existing baselines in terms of false discovery control, precision, and recall, highlighting the trade-offs between suppressing false detections and retaining true objects.

Critical strengths

The paper effectively integrates e-value-based FDR control into OVD through image-block calibration. By accommodating score-dependent candidate counts and within-image dependence, it provides a principled approach to suppressing false detections without retraining the detector.

The theoretical analysis is well developed, with explicit assumptions and detailed derivations of the FDR guarantees and release conditions. These results support the statistical foundation of the method and explain how evidence strength, candidate count, and rank influence which detections are retained.

Critical weaknesses

The writing lacks clarity and makes the paper difficult to follow. Although the task is formally defined, its practical motivation and relationship to prior work are insufficiently explained. This makes it difficult to understand the intended application and assess the contribution beyond existing FDR methods.

The results in Table 2 and Figure 3 do not show consistent improvements over Monotone CRC, leaving the practical benefits of the proposed method unclear.

The prompt-bank experiment raises concerns about scalability. With 40 calibration images, recall drops from 43.62% to 2.04% as the prompt count increases from 80 to 1,280 (Figure 6). This substantial loss of useful detections limits the demonstrated applicability to larger prompt banks.

The FDR guarantee assumes a fixed α, but the adaptive procedure in Section 4.1 selects α using the calibration data. The paper does not establish whether FDR remains controlled after this selection.

Questions for authors

Could the authors clarify the additional benefit of the proposed method over Monotone CRC, given the results in Table 2 and Figure 3?

1 comment
Reviewer_75RaOpenReview ↗

Reply to authors

I thank the authors for their efforts to address my questions. They have provided some new results to support the performance. However, the reply for contribution is still unconvincing, it is recommended the authors to summarize any technical or theoretical contributions over previous approaches. Besides, my largest concern is still the paper quality, which is not improved in this rebuttal.

Reviewer_CC1kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Block Calibration, a postprocessing framework for deciding which detections from a frozen open-vocabulary detector should be released. Rather than pooling a variable number of proposals across queries, the method treats each image and its associated queries as an exchangeable block. It averages null evidence within each query and then across queries in the image, constructs block-normalized evidence using Eqs. (2)–(3), and applies e-BH within each query through Eq. (4). The paper proves control of the expected mean-family FDP defined in Eq. (1), introduces a maximum-reference variant for the within-image pooled FDP in Eq. (6), and evaluates the method on GroundingDINO and YOLO-World across COCO, LVIS, and Open Images.

Critical strengths
  1. Well-motivated calibration unit. The paper addresses an important reliability problem in open-vocabulary detection. Its counterexample effectively shows why candidate-level pooling can overweight queries with many proposals.

  2. Clean theoretical result. The main argument appears correct. The bound in Eq. (5), together with block exchangeability and the inequality U0≥B0U_0 \geq B_0, accommodates random candidate counts and arbitrary dependence within an image.

  3. Useful characterization of power. The release condition in Eq. (7) provides useful insight into how evidence strength, family size, and rank affect selection, and helps explain why proposal filtering can improve recall.

Critical weaknesses
  1. Potential gap between the primary guarantee and deployment-facing error. Eq. (1) averages FDP equally across query families and defines FDP as zero for empty releases. As a result, mean-family FDP may not always closely reflect the false fraction among all released detections, depending on how errors and release counts are distributed across queries. The maximum-reference variant associated with Eq. (6) appears more directly related to an image's pooled output. It would be helpful to include this variant more systematically in the main cross-dataset evaluation.

  2. Clarification of cluster-level versus box-level control. The cluster extension defines a cluster as correct when at least one member box is correct, while the description suggests that its members may be released together. In clusters containing both correct and incorrect boxes, cluster-level FDP may therefore differ from box-level FDP. The reported singleton statistics for the primary COCO configuration suggest that this distinction may often be small, but reporting box-level FDP, member-box precision, and cluster-size statistics across the main settings would clarify its practical importance.

  3. Some operating-point comparisons are retrospective. Table 1 selects method-specific nominal levels from the measured precision–recall curves, while Table 15 forms matched comparisons based on observed test precision. These analyses are useful for describing the empirical frontier, and the paper appropriately notes their retrospective nature. However, they may not represent operating points that could be selected before deployment without access to test labels. It would strengthen the empirical conclusions to emphasize results based on fit- or calibration-only selection and to clearly separate these from test-matched descriptive comparisons.

Questions for authors

Several experimental splits reuse or repartition overlapping image pools, while Table 15 reports normal-approximation confidence intervals over matched comparisons. Could the authors clarify how dependence among repeated images, splits, and operating points is handled in these intervals? An image-level bootstrap over unique images or an evaluation on independent holdout pools may provide a more interpretable estimate of uncertainty.

Reviewer_vmJ1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the selection of final detection results in open-vocabulary object detection, aiming to maximize the retention of true objects while controlling the false discovery rate (FDR). Addressing challenges such as the variable number of candidate boxes per query, interdependencies among these boxes, and the potential misalignment between a global calibration approach (mixing all boxes together) and the objective of controlling the FDR on a per-query basis, the paper proposes a patch-based calibration method. By treating each image, along with its associated queries, as a distinct unit and incorporating the e-BH multiple-testing procedure, this approach provides per-query average FDR guarantees, under the assumption that the image patches are exchangeable.

Critical strengths

The paper’s strength lies in addressing a problem with clear practical significance, particularly for applications such as visual search and automated tagging, where the suppression of false detections is essential. The proposed method aligns calibration units with target risks, allowing the number of candidates to depend on both detection scores and intra-image dependencies, all without requiring the detector to be retrained. The approach is elegantly constructed and grounded in a solid theoretical framework. Through experiments covering various detectors and datasets, the authors report metrics including error rates, precision, recall, and the number of output candidates. They also examine the impact of evidential learning, candidate filtering, and prompt scale, providing a comprehensive look at the trade-off between error control and target retention. Furthermore, at certain high-precision operating points, the method achieves superior recall compared to several baselines.

Critical weaknesses

The primary limitation of this paper is that the scope of its technical innovations and performance advantages requires clearer definition. The core methodology builds upon existing e-BH procedures and exchangeability arguments; the "exact release condition" is essentially an algebraic reformulation of the screening rules, and the clustering extension follows a similar proof structure. Consequently, a more in-depth comparison with closely related grouping, multi-layer, and conformal multiple-testing methods is necessary. The experiments fail to demonstrate consistent superiority over strong baselines: in comparisons involving similar precision levels, the proposed method underperforms relative to monotonic CRC and calibrated score thresholds, while exhibiting low recall on certain datasets. Furthermore, some operating points and pairings appear to rely on test-set results; it remains unclear whether the reported normal-approximation confidence intervals account for overlapping partitions, repeated operating points, or the issue of multiple comparisons. It is recommended that operating points be selected using an independent validation set, with fair comparisons conducted on a separate test set. Statistical inference should also account for dependencies between image patches. Additionally, the evaluation should be supplemented with conditional risk (given the output queries), category-specific risk, and tail-risk metrics, alongside standard one-to-one matching and detection metrics across multiple IoU thresholds—thereby ensuring that average error rates and single-threshold localization criteria do not obscure underlying issues with output quality.

Questions for authors

The authors are requested to clarify the following points: (1) Compared to the most closely related existing statistical calibration methods, what are the specific new contributions of chunk-based normalization that cannot be directly derived from prior work? (2) When the operating point is selected completely independently of the test labels, can this method consistently outperform monotonic CRC, calibration score thresholding, and FP-CP under the same risk or precision constraints? (3) In scenarios involving low recall—or a significant drop in recall as the prompt scale increases—is the primary bottleneck related to candidate coverage, the ability to discriminate evidence, or the inherent conservatism of the calibration rules?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.