acceptodds
Under review as a conference paper at ICLR 2027

Counting Beyond Instances: BunchCount for Group-Individual Object Counting

Abstract

Visual counting is commonly formulated at the instance level, aiming to estimate how many objects of a queried category appear in an image. However, real-world counting often involves higher-level semantic units formed by multiple instances, such as a bunch of grapes, a stack of plates, or a pair of shoes. This exposes a key limitation of existing counting formulations, which mainly focus on what to count, while largely overlooking at which semantic unit to count. We introduce Group-Individual Object Counting (GIC), a new setting that requires models to count both individual objects and semantic groups within a unified framework. To support this new task, we present BunchCount, a real-world benchmark with 1,330 images, 89,254 individual annotations, and 11,065 group annotations. BunchCount provides paired individual-group annotations within the same image and explicitly records containment relations between each group and its constituent individuals. Experiments on BunchCount show that current advanced counting models perform well on individual instances but fail to count semantic groups more accurately. To mitigate semantic granularity conflict, we propose a counting-unit guided relational counting framework, which exploits group-individual containment relations to regularize cross-granularity representations during training. Our method substantially improves group-level counting while better preserving individual-level counting ability, establishing a strong baseline for counting beyond instances.

Primary area:
datasets and benchmarks
Keywords:
Visual Counting, Counting Unit, Semantic Group, Group-Individual Object Counting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_f6sCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Group-Individual Object Counting (GIC), a task that requires models to count either individual objects or semantic groups according to the specified counting unit. To facilitate research on this task, the authors construct BunchCount, a benchmark containing 1,330 images with paired individual- and group-level annotations and explicit group-individual containment relations.

The authors further propose GICount, a framework built upon CountGD that incorporates counting-unit prompt decomposition and relation-based supervision. Specifically, it constructs positive and negative sets of individual queries based on annotated group membership, aggregates their representations using a Transformer, and aligns them with group-structure semantics through a contrastive objective. Experiments demonstrate improved group-level counting performance while maintaining competitive individual-level counting accuracy.

Critical strengths

The paper identifies an interesting aspect of visual counting by distinguishing between individual instances and semantic groups. BunchCount provides paired annotations and explicit containment relations, making it a potentially valuable benchmark for investigating counting across different semantic units.

Critical weaknesses

The proposed method largely extends CountGD, which does not appear to introduce a substantially new mechanism for group-individual counting.

The reported improvement in group counting is overwhelmingly concentrated in the fixed-cardinality subset, where the MAE decreases from 58.22 to 8.11. Improvements in the remaining group structures are relatively modest, and performance even slightly deteriorates for spatial aggregation.

Questions for authors

Why is the overall improvement primarily driven by the fixed-cardinality subset?

Can the method generalize to previously unseen group structures, rather than only unseen object categories?

2 comments

Thank you for recognizing the value of the task and benchmark. Below, we address the methodological contribution, the gains on fixed cardinality, and generalization to unseen organizational forms.

Response to Weakness 1: Methodological contribution

GICount adds explicit group–individual relation supervision to the CountGD backbone. We choose CountGD for its strong counting performance. We introduce a relation module that aligns instance combinations with counting-unit semantics using annotated group-individual containment relations (§5.2–5.3 and Figure 4). The ablations in Table 2 directly evaluate this design.

MethodInd. Test MAE ↓Ind. Test RMSE ↓Grp. Test MAE ↓Grp. Test RMSE ↓
w/o relation modeling (CountGD)13.5438.689.2128.30
w/o counting-unit prompt18.3657.391.936.79
full (Ours)12.2028.891.855.49

Relation supervision strengthens group recognition, while the counting-unit prompt enables this supervision to improve both individual and group counting. The full model adds only 1.1% parameters, and the relation module is used only during training (§5 and Table 2; Appendix D).

Response to Weakness 2: & Question 1: Gains on fixed cardinality

The aggregate gain is dominated by fixed-cardinality groups, where jointly trained CountGD incurs exceptionally large errors. Fixed-cardinality groups form a whole through a conventional or functional number of constituent instances, such as a pair of shoes or two lenses forming a pair of sunglasses (Appendix A.2). Such groups are defined by semantic pairings or combinations, rather than by spatial proximity or a shared boundary. Existing counting methods do not explicitly model these group-unit semantics, making fixed cardinality a challenging test of counting-unit understanding. Appendix Table 4 reports results for the four structures. We add GroundingREC results below; each entry reports group-level Test MAE / RMSE:

MethodPhysical Connection (39 images)Spatial Aggregation (171 images)Fixed Cardinality (45 images)Structural Arrangement (55 images)
CountGD1.64 / 4.110.64 / 1.9858.22 / 74.051.11 / 1.93
GroundingREC6.74 / 14.580.89 / 3.1512.42 / 23.132.53 / 3.74
GICount1.18 / 1.950.74 / 2.248.11 / 13.550.62 / 1.18

Both baselines incur their largest errors on fixed cardinality. Supervision from group-individual relations teaches GICount how constituent instances form counting units, and its substantial improvement on this subset reflects a better semantic understanding of group counting units. On the other three group structures, GICount also achieves substantial relative improvements or remains competitive, demonstrating the value of relation modeling across different group structures.

Response to Question 2: Generalization to unseen organizational forms

GICount’s gains extend to organizational forms excluded from BunchCount training. We hold out 9 organizational forms: strings and bracelets (physical connection); baskets, bowls, and buckets (spatial aggregation); sets (fixed cardinality); and rows, columns, and lines (structural arrangement). This yields 1,032 training images and 298 test images; both models use the original initialization and train for 20 epochs.

MethodInd. Test MAE ↓Ind. Test RMSE ↓Grp. Test MAE ↓Grp. Test RMSE ↓
CountGD8.7634.7412.3432.81
GICount8.2131.099.1519.61

GICount improves both group and individual counting on organizational forms unseen during training, showing that the group representations learned through relation supervision can transfer to new organizational forms.

Thank you again for your constructive feedback and for recognizing the value of our work. We are happy to address any remaining questions or concerns.

Thank you again for your comments. Our rebuttal presents further analyses of relation modeling and the gains on fixed-cardinality groups, together with a new structure-holdout experiment evaluating generalization to unseen group structures. As the discussion period is nearing its end, we would appreciate your assessment of whether these results address your concerns. Please let us know if any points require further clarification.

Reviewer_Y62mOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new task called "Group-Individual Counting" (GIC). To solve this new task, this manuscript also build a real-world benchmark, i.e., BunchCount, which annotates the inclusion relationships between groups and individuals.

Critical strengths
  1. The task definition is novel and addresses a genuine gap in existing benchmarks. This formulation is intuitive and aligns well with real-world applications.
Critical weaknesses
  1. The definition of a "semantic group" is ambiguous, and annotation consistency is not quantified. The paper also acknowledges that a group is "primarily a semantic and perceptual unit rather than one determined by a single geometric rule" in Appendix A.1. Several boundary cases remain unspecified. For example, how should a fruit basket containing both apples and oranges be labeled, one group or two per-category groups? Similarly, no operational threshold is given for how close objects must be to form "one pile" versus "two piles," or whether a group remains valid under partial occlusion, and no inter-annotator agreement (e.g., IoU or Cohen's kappa) is reported. Since determining group boundaries is the core of the GIC task, this gap directly affects the credibility of the benchmark and of the conclusions drawn from it.
Questions for authors

See weaknesses

4 comments

Thank you for recognizing the significance of the task. We address your concerns in four parts: mixed-category grouping, spatial grouping and distance thresholds, partial occlusion, and annotation consistency.

Mixed categories. Groups are annotated according to the target category and counting unit specified by the query. For apples and oranges mixed in a basket, “basket of fruits” includes both under the category of fruit and counts the entire basket as one group. If the two types of fruit form two separate piles within the same basket, they can also be queried and annotated as “pile of apples” and “pile of oranges,” respectively. For category-specific queries, only instances of the specified category are eligible group members. Both the target category and counting unit are fixed for each annotation.

Spatial distance and grouping. Appendices A.1 and A.2 define groups using visible organizational cues, including aggregation, separation, connection, arrangement, and containers. Local clustering within a pile and separation between piles describe the overall spatial organization; they do not imply a fixed distance threshold between instances. In Appendix Figure 1, piles of coffee beans are distinguished by their clustered shapes and separation; crows in the same row form a group through their arrangement even when spaced far apart; blueberries in neighboring boxes form groups defined by their respective containers.

It is noted that a distance threshold alone cannot reliably recover these groups. We directly test whether thresholding Euclidean distances between instances can recover the grouping: the threshold must connect the constituent instances within each group without merging neighboring groups. We check whether such a threshold exists separately for each pair of neighboring groups, with the following results:

Group StructureValid Neighboring Group PairsPercentage Not Recoverable by a Distance Threshold
Physical Connection1,19939.62%
Spatial Aggregation1,55925.72%
Fixed Cardinality2,17438.04%
Structural Arrangement1,57024.71%

For these pairs, even a separately chosen threshold cannot maintain both within-group connectivity and between-group separation. This shows that group boundaries cannot be determined by a uniform distance threshold between instances and require recognition of the actual organizational relations.

Partial occlusion. In Appendix Figure 1, the stack of bowls and the buckets of apples in the first row remain identifiable as groups despite partial occlusion, based on the overall stack structure and container outlines, respectively. Following Appendices A.3 and A.4, individual point annotations mark only visible instances and each group box covers its visible constituent instances. Groups without an explicit enclosing boundary must contain at least two visible instances. When only one instance is visible, a clear container boundary or a recognizable whole-object structure supports the group identity. Candidate images are excluded during screening if occlusion prevents individual identification or if group boundaries or group-individual containment relations cannot be determined.

Annotation consistency. After the six annotators completed the initial annotations, we conducted three rounds of correction and consistency checking (§4.1 and Appendix A.4). The first round identified discrepancies in group counts in 73 images. Subsequent cross-checks covered text annotations, missing or duplicate points and boxes, and group-individual containment relations. Disputed groupings were reviewed by two other annotators under the same guidelines and checked again after correction. After the final round, 58 images still had at least one annotation disagreement over the grouping, accounting for less than 5% of all 1,388 candidate images. These images were excluded, leaving the final set of 1,330 images with unambiguous group annotations.

We appreciate your encouraging assessment and thoughtful suggestions. We look forward to continuing the discussion and addressing any points you would like us to clarify.

Reviewer_Y62mOpenReview ↗

Follow-up comments

Thanks the author for your clarification.

However, several issues remain.

  1. How was the query for each image selected? The response defines groups relative to a query-specified category and counting unit, but does not explain how queries were chosen or whether the same image admits multiple equally valid groupings.

  2. The existence of these removed 58 images is itself evidence of the task's ambiguity: if group boundaries were uniquely determined, six annotators would not have remained in disagreement after three rounds of correction. Excluding them therefore removes the evidence of ambiguity rather than resolving it. This directly bears on the benchmark's credibility.

Overall, the definition of the task itself remains substantially ambiguous, which directly affects the credibility of the benchmark and its conclusions.

Thank you for your follow-up questions.

Response to Question 1: Query selection and grouping consistency

An image can support multiple valid counting targets. For example, the rows and columns formed by the same set of regularly arranged objects provide different group-level counting targets. For each image, expert annotators select an object category and a group description grounded in the image content and provide annotations for that target.

Annotators select the group description based on the object category and visible organization in the image, following the definitions of group structures and annotation guidelines in Appendix Sections A.1–A.2. The descriptions are then cross-checked against the visible organization (Appendix A.4). REC-8K likewise constructs referring expressions from visible object attributes, with multiple expressions for the same image specifying different counting targets [1]. A query can therefore specify its target precisely without being the image's only valid description.

For a fixed query, we establish grouping consistency through shared annotation guidelines and cross-checking. Annotators use visible cues such as connection, arrangement, aggregation, and container boundaries to determine group boundaries and group–individual containment relations (Appendix Sections A.1 and A.4). Disputed groupings are reviewed by two other annotators under the same guidelines and checked again after correction.

Response to Question 2: Inferences drawn from annotation disagreements

We disagree with the inference from annotation disagreements on particular candidate images to a “substantially ambiguous” task definition. As reported in Appendix A.4 and our first response, disagreements over group annotations persisted for 58 candidate images after review. These disagreements concerned whether a group was discernible(15 images), how adjacent groups should be separated(14 images), and which individual objects belonged to each group(29 images). The reported total combines these distinct forms of annotation disagreement, which cannot all be equated with multiple equally valid groupings under a fixed query.

Excluding samples with annotation ambiguity is a common practice for maintaining the quality of benchmark datasets. In salient object detection, VizWiz-SO annotators selected the child, the cup, or no single foreground object in the same clear image (Supplementary §A.3). The benchmark excludes candidates whose foreground remains ambiguous after review [2]. In human–object interaction detection, CrossHOI-Bench shows a clear still image consistent with either boarding or exiting a plane (Figure 1), and its evaluation accepts such alternatives [3]. These examples show that semantic ambiguity in particular images can coexist with a well-defined task and evaluation protocol.

BunchCount explicitly defines semantic groups and their annotation rules. A group comprises individual objects of the same category whose recognizable organization allows them to be counted as a single semantic unit (Appendix Sections A.1–A.2). Annotations include group bounding boxes, points for individual objects, and explicit group–individual containment relations. Each individual object is associated with at most one group, and objects may remain ungrouped (Appendix A.4).

Excluding these samples protects the reliability of supervision and evaluation. It does not undermine the benchmark’s credibility. Forcing a grouping that remains disputed after review to serve as ground truth would make training targets and evaluation scores depend on an ambiguous annotation decision. BunchCount provides both group counts and group–individual containment relations, so disagreement at either level can affect the reliability of these annotations.

These 58 images constitute less than 5% of the candidate images and were excluded during annotation review. We hope these explanations clarify how the query and annotation protocol establish consistent counting targets in BunchCount.

Thank you again for your feedback. We would be happy to address any further questions.

[1] Referring Expression Counting. CVPR 2024.

[2] Salient Object Detection for Images Taken by People With Vision Impairments. WACV 2024. Supplementary §A.3.

[3] CrossHOI-Bench: A Unified Benchmark for HOI Evaluation across Vision-Language Models and HOI-Specific Methods. CVPR 2026.

Thanks for your response.

However, I am still concerned about the definition of your proposed task. I keep my rate.

Reviewer_Zra3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Group-Individual Object Counting, a new task that requires counting either individual objects or semantic groups. The authors construct BunchCount and propose GICount, which uses group-individual relations to improve counting at both levels. The results show promising improvements on the test set. However, some experimental results are difficult to explain, and the evaluation does not fully establish the advantage of the proposed method.

Critical strengths

The paper introduces a new counting task with practical value. The paired individual- and group-level annotations provide a useful benchmark for studying different counting units. GICount also achieves strong group-level results on the test set, showing the potential of relational supervision.

Critical weaknesses
  1. The group annotations are represented by bounding boxes, and the proposed method is built on a detection-based counting framework. Given the annotation format and the distribution of group counts (at most ~50), it is unclear whether this task is better formulated as counting or as group detection. The paper should compare with some stronger detection-based baselines to justify the need for a dedicated counting framework.

  2. The results in Table 1 are difficult to explain. Some baselines show a large performance drop from validation to test. For example, GroundingREC's group MAE increases from 0.97 to 3.59, while CountGD's increases from 1.77 to 9.21. In contrast, GICount's group MAE changes from 2.18 to 1.85. The authors should analyze these differences and clarify whether they result from differences in group-count distributions, category difficulty, or model generalization.

  3. GICount does not outperform some baselines on the validation set, yet show better group-level performance on the test set. For example, its group MAE is 2.18, compared with 0.97 for GroundingREC and 1.77 for CountGD on validation, but it achieves 1.85 on test, compared with 3.59 and 9.21, respectively. Since CountGD is also the backbone of GICount, this result needs further explanation. More analysis is needed to establish that the improvement comes from the proposed relation modeling rather than differences in generalization or model selection.

Questions for authors

See three points in weakness please.

6 comments

Thank you for recognizing the value of the new task and the paired annotations. We have added a detection baseline and analyses by group structure. Below, we address the task formulation, performance changes across splits, and the roles of relation modeling and model selection.

Response to Comment 1: Group annotations, counting formulation, and detection baselines

Group counting addresses practical needs, with detection as a natural solution. Early studies on group counting, including grape-bunch detection and counting in agriculture (Íñiguez et al., 2024), remained confined to specific domains. GIC extends this problem to a class-agnostic setting, studying whether one model can count individuals or semantic groups according to the prompt (§3).

Group-box annotations do not imply full-box localization supervision. To study this task with a strong and widely used counting framework, we build on the detection-based counting model CountGD. We retain group bounding boxes to record each group's spatial extent and provide visual evidence of the group and its constituent instances. During training, all counting models only use group-box centers as group-counting supervision (§4.1 and Appendices A.4 and B.1).

The group-count range reported in §4.2 and Figure 3 is 1–253, with a mean of 8.32 and a median of 4 groups per image. The final histogram bin in Figure 3 denotes counts above 50. We believe that neither the count range nor bounding-box annotations alone determine whether a task is a counting task: CARPK provides full vehicle boxes and compares counting methods using MAE/RMSE; CountBench contains only 2–10 objects per image and is also used to evaluate counting ability.

We added GroundingDINO detection baselines under joint and group-only training, using center-point and full-box supervision, respectively. The following table extends Table 1. Each entry reports MAE / RMSE.

MethodInd. Val ↓Grp. Val ↓Ind. Test ↓Grp. Test ↓
GroundingDINO [T]16.07 / 42.712.64 / 9.3527.14 / 76.0311.06 / 35.07
CountGD [T+V] (Table 1)13.05 / 34.051.77 / 7.7913.54 / 38.689.21 / 28.30
GICount [T+V] (Table 1)13.07 / 31.982.18 / 9.3112.20 / 28.891.85 / 5.49

The counting models outperform the general-purpose detector, and GICount further improves test performance for both counting units. The following table extends Table 3 with the group-only and joint detection baselines.

MethodTrain SettingInd. Test MAE ↓Ind. Test RMSE ↓Grp. Test MAE ↓Grp. Test RMSE ↓
GroundingDINOGrp.-only——6.3013.45
CountGDGrp.-only73.95115.761.383.98
GroundingDINOJoint27.1476.0311.0635.07
CountGDJoint13.5438.689.2128.30
GICountJoint12.2028.891.855.49

Even with full-box supervision, the group-only detector underperforms the group-only counting model. These results demonstrate that straightforward detector adaptation is insufficient to address GIC effectively. The challenge extends beyond locating objects to selecting the intended semantic counting unit and supporting both individual and group counting within one model. This motivates dedicated evaluation and modeling of GIC.

Response to Comment 2: Sources of the validation-to-test performance changes

Differences in category and group-structure distributions help explain the validation–test discrepancy. BunchCount is split into training, validation, and test sets by object category. Categories overlapping with FSC-147 retain their original split assignments to prevent category leakage. The splits therefore differ in category composition, group counts, and structure proportions. The four structures (physical connection, spatial aggregation, fixed cardinality and structural arrangement) involve different grouping cues, which present different challenges for existing methods. The split rationale and statistics are provided in §4.2 and Appendix A.5.

Appendix Table 4 reports results by group structure. We extend the structure-wise test analysis with GroundingREC below; each entry reports group-level MAE / RMSE.

MethodPhysical Connection (39)Spatial Aggregation (171)Fixed Cardinality (45)Structural Arrangement (55)
CountGD1.64 / 4.110.64 / 1.9858.22 / 74.051.11 / 1.93
GroundingREC6.74 / 14.580.89 / 3.1512.42 / 23.132.53 / 3.74
GICount1.18 / 1.950.74 / 2.248.11 / 13.550.62 / 1.18

GICount’s stronger test performance comes mainly from reducing errors on the structures where the baselines struggle most. CountGD, GroundingREC, and GICount all have low group-counting MAEs on the validation set. On the test set, the increases in error for CountGD and GroundingREC are concentrated in more challenging group structures, particularly fixed cardinality and physical connection. GICount's improvements on these structures account for its stronger test performance.

We further assess generalization to unseen organizational forms in the holdout experiment presented in our response to Comment 3 below.

Response to Comment 3: Validation rankings, the effect of relation modeling, and model selection

A few counting-unit switching failures increase validation MAE. Our per-image analysis shows that, on a small number of validation samples, GICount predicts counts close to the individual counts even when given group prompts, failing to switch counting units correctly. These failures produce large group-counting errors and affect its overall validation performance.

The proposed relation-modeling components distinguish GICount from CountGD and jointly improve overall counting performance. The counting-unit prompt, relation aggregation, and associated supervision together constitute our relation-modeling method. Removing this mechanism leaves only the individual- and group-level detection losses, reducing the model to jointly trained CountGD, i.e., w/o relation modeling in Table 2. Thus, the comparison between CountGD and the full GICount model directly tests the effect of relation modeling. We reproduce Appendix Table 4 below. Each entry reports group-level Test MAE / RMSE. Relation modeling improves group counting on three of the four structures.

ModelPhysical Connection (39)Spatial Aggregation (171)Fixed Cardinality (45)Structural Arrangement (55)
w/o relation modeling (CountGD)1.64 / 4.110.64 / 1.9858.22 / 74.051.11 / 1.93
full (Ours)1.18 / 1.950.74 / 2.248.11 / 13.550.62 / 1.18

Model selection uses validation results only, and the test gains persist across nearby checkpoints. Following Appendix Eq. (2), we normalize the validation MAE of each unit by its lowest value across candidate epochs and take their equally weighted average. This score selects epoch 17 from GICount's 30 candidate epochs for the results in Table 1. The following table shows validation results for 10 consecutive epochs around the selected epoch.

EpochInd. ValGrp. ValInd. TestGrp. TestValidation Selection Score
1215.58 / 36.062.42 / 9.43——1.159572
1314.03 / 34.122.33 / 9.31——1.079956
1414.85 / 35.142.50 / 9.48——1.150214
1513.37 / 32.292.34 / 9.3511.85 / 27.521.82 / 5.291.057954
1613.58 / 33.042.23 / 9.37——1.038285
17 (selected)13.07 / 31.982.18 / 9.3112.20 / 28.891.85 / 5.491.007716
1814.33 / 33.972.15 / 9.37——1.048403
1913.48 / 33.452.28 / 9.38——1.046830
2015.11 / 35.392.21 / 9.3813.25 / 30.031.75 / 5.231.092196
2114.13 / 33.592.21 / 9.41——1.054690

GICount generalizes to organizational forms unseen during training. We hold out 9 organizational forms: strings and bracelets (physical connection); baskets, bowls, and buckets (spatial aggregation); sets (fixed cardinality); and rows, columns, and lines (structural arrangement). This yields 1,032 training images and 298 test images; both models use the original initialization and train for 20 epochs. GICount improves both group and individual counting, demonstrating the transferability of the group representations learned through relation supervision.

MethodInd. Test MAE ↓Ind. Test RMSE ↓Grp. Test MAE ↓Grp. Test RMSE ↓
CountGD8.7634.7412.3432.81
GICount8.2131.099.1519.61

Thank you for your constructive feedback. Please let us know if any concerns remain. We would be happy to address them and provide further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.