acceptodds
Withdrawn from ICLR 2027

Counting the Uncounted: Correcting Occlusion-caused Bias in Crowd Counting Benchmarks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-based crowd counting has advanced considerably, yet its real-world reliability remains limited by occlusion, which existing methods treat as undifferentiated noise rather than as a structured scene element. We challenge a foundational assumption of current benchmarks: that ground-truth (GT) annotations reflect true crowd size. By auditing four widely used datasets (ShanghaiTech A, UCF-QNRF, JHU-Crowd++, and NWPU-Crowd), we introduce a three-class occlusion taxonomy: Atmospheric (AO), Barrier (BO), and Embedded Occlusions (EO). We identify EOs as the most consequential yet overlooked category. EOs are small, non-transparent objects embedded within the crowd, such as umbrellas, picket signs, and flags. Occlusions appear in 40.73% of images, and 78.46% of these involve EOs, yet no annotation protocol accounts for the individuals concealed behind them. We correct the GT through occluder-specific annotation and local density inference with a new annotation tool, Occlusion Density Fusion (ODFlow). The correction shows that existing GT undercounts EO-affected images by 6.3% on average, so published MAE understates real-world error by about 22 counts per image. Beyond data correction, we propose ODFNet, an occlusion-aware counting model. Its Density Reconstruction and Count Fusion (DRCF) module reconstructs human and EO density maps separately, then fuses them while removing redundancy introduced by overlapping EOs. Validation on both synthetic crowd scenarios and real crowd images shows that ODFNet recovers the ground-truth count regardless of EO type, number, or spatial pattern, whereas existing methods drift substantially. We release corrected GT for all four datasets, a comprehensive benchmark re-evaluation, a newly curated occlusion-aware dataset (ODF-A), and ODFNet.

Primary area:
datasets and benchmarks
Keywords:
Crowd Counting, Embedded Occlusions, Ground-Truth Correction, Occlusion Taxonomy, Crowd Datasets, Annotation Quality, Computer Vision

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_1spCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the potential undercounting problem caused by embedded occluders in crowd-counting datasets. The authors categorize occlusions into three types: atmospheric, barrier, and embedded occlusions. The proposed framework separately predicts human and occluder density maps, estimates the density of people potentially hidden behind occluders, and integrates these representations to obtain the final crowd count. The problem is interesting, but several aspects remain unclear.

Critical strengths

The paper provides a comprehensive solution including problem identification, model development, corrected ground-truth annotations, and the construction of a dedicated occlusion-aware dataset.

Critical weaknesses
  1. If the original benchmark annotations count only identifiable or visible heads, introducing estimates for fully occluded people changes the prediction target rather than correcting annotation errors. The authors should more clearly justify why such hidden individuals should be included in the ground-truth count and explain how comparisons with existing crowd-counting methods remain fair under this modified target definition.

  2. The proposed model primarily considers umbrellas and picket signs as embedded occluders. However, other common occluders in crowded scenes, such as flags, banners, boards, are not evaluated.

  3. The presence of an occluder does not necessarily imply that people are hidden behind it. It is therefore unclear whether the model tends to predict hidden people whenever such occluders are present. For example, would an image containing a sign but no person behind it lead to a nonzero hidden-human density or an overestimated crowd count? Additional analysis or controlled experiments would help clarify this behavior.

  4. The computational cost of the proposed method is not reported. In addition, it is unclear whether the method is restricted to (512*512) inputs.

Questions for authors

Minor: The roles of S_b and C_b in reconstructing the density map should be explained more clearly. The manuscript should also specify how the different predicted density maps are fused to produce the final output.

In Figure 2, the values 228.4 and 24.4 are reported together with a final count of 263.7. However, the relationship among these values is unclear, since the first two values do not directly sum to the reported final count. The authors should explicitly explain how the final count is computed.

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

This paper investigates occlusion-induced annotation bias in crowd counting benchmarks. The authors introduce a taxonomy of atmospheric, barrier, and embedded occlusions, and propose ODFlow to generate corrected ground-truth annotations for four existing datasets. They further construct an occlusion-aware dataset and propose ODFNet, which separately models visible-human and occlusion-related densities to estimate concealed individuals. Experiments on corrected benchmarks and synthetic occlusion scenarios demonstrate improved counting performance under the proposed evaluation protocol.

Critical strengths
  1. The paper raises a meaningful question about the reliability of existing crowd counting annotations under severe occlusion. Investigating potentially overlooked individuals could contribute to more accurate and realistic evaluation benchmarks.
  2. The work combines dataset auditing, occlusion-aware annotation, and a corresponding counting framework. The synthetic occlusion experiments provide preliminary evidence that explicit occlusion modeling can improve counting robustness.
Critical weaknesses
  1. The reliability of the corrected ground truth is not sufficiently established. The proposed EO-GT estimates concealed individuals through occluder annotations and local density inference, but lacks independent verification of the actual number of hidden people. This raises concerns about whether the proposed annotations genuinely correct existing benchmark errors or simply introduce estimated counts under a different annotation protocol.
  2. The experimental validation is insufficient to support the main claims. ODFNet is primarily evaluated against the authors' corrected annotations, making it difficult to distinguish genuine counting improvements from adaptation to the proposed annotation scheme. Moreover, the synthetic occlusion experiments are based on a single original scene, and the contribution of the proposed DRCF module is not adequately isolated through controlled ablations.
  3. The manuscript also requires improvements in organization and presentation. Section 2 mixes dataset construction with network architecture, and some table references are inconsistent (e.g., Table 6 is cited instead of Table 5 in Section 3.1). These issues should be corrected to improve clarity and readability.
Questions for authors
  1. How are the numbers of fully occluded individuals verified when constructing EO-GT? Can the authors provide independent evidence that the corrected annotations better reflect actual crowd sizes?
  2. How do the authors distinguish genuine annotation errors from differences in annotation protocols? Do the original benchmarks explicitly require counting fully invisible individuals?
  3. Can the authors provide more comprehensive evaluations using independently verified ground truth and additional synthetic scenes? How is potential overlap between ODF-A and the original benchmark datasets handled during training and evaluation?
  4. Can the authors provide controlled ablation studies to clarify the independent contribution of DRCF compared with simpler occlusion-aware count fusion strategies?
Reviewer_HhMnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an occlusion-aware crowd counting framework that addresses potential undercounting caused by occluded individuals in existing datasets. The authors introduce a three-class occlusion taxonomy, an annotation correction tool called ODFlow, and a counting network called ODFNet that separately reconstructs human and occluder density maps. Experiments on four datasets suggest that existing annotations may underestimate crowd sizes and demonstrate the effectiveness of the proposed framework in handling occlusions.

Critical strengths

The paper investigates the potential undercounting problem caused by occlusions in existing crowd counting datasets, which is an interesting and practically relevant issue.

Critical weaknesses
  1. It is unclear how fully occluded individuals can be assigned accurate ground-truth annotations when they are not visible in the image. The reliability of these corrected annotations is questionable.

  2. The distinction between the proposed task and conventional multi-object counting is unclear. Since occluders can be treated as additional object categories, the authors should explain the fundamental differences and challenges.

  3. The novelty of the proposed method is limited. The proposed density reconstruction and fusion strategies appear closely related to existing crowd counting and multi-object counting approaches.

Questions for authors
  1. How can the authors guarantee the correctness of ground-truth annotations for individuals who are completely invisible due to occlusion?

  2. How does ODFlow infer the number of invisible individuals, and what evidence supports the accuracy of these inferred annotations?

  3. What are the fundamental differences between the proposed occlusion-aware crowd counting task and conventional multi-object counting with additional object categories?

  4. Why is a specialized framework necessary instead of directly applying existing multi-object counting methods?

  5. What are the key technical innovations of ODFNet compared with existing density estimation and feature fusion methods?

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

Related papers

Discussion (0)

Sign in to comment.