PaperlayerView exact revision on arXiv

Source-linked AI summary

Counting Beyond Instances: A Benchmark for Group-Individual Object Counting

Rui Wang, Junyi Huang, Jiahui Li, Qiao Yu, Yixue Hao, Long Hu, Baoru Huang

arXiv:2609.04716v1cs.CV

TL;DR

Existing counting methods largely select what to count but not whether the unit is an individual object or a semantic group. The paper introduces GIC and BunchCount, then proposes GICount, whose relational framework improves group counting while preserving individual counting ability. On BunchCount, GICount substantially advances group-level counting and retains strong individual-level performance.

  • Problem

    Existing counting paradigms mainly use individual instances as counting units and do not evaluate switching between individual- and group-level counting for the same category and image.

  • Method

    GICount uses counting-unit prompts and group-individual containment relations to regularize representations across individual and group counting.

  • Results

    GICount improves group-level counting while preserving individual-level counting ability, achieving group-test MAE 1.85 and RMSE 5.49 versus GroundingREC.

  • Takeaways & Limitations

    GIC and BunchCount establish a benchmark and strong baseline for counting semantic units beyond individual instances.

Abstract

from arXiv · show

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.

Introduction

The paper extends visual counting from selecting target categories to selecting semantic counting units, enabling models to count either individual objects or groups. It introduces BunchCount and GICount to evaluate and address the difficulty of unified cross-granularity counting.

  • Existing counting paradigms mainly count individual instances, despite advances in specifying target categories, exemplars, and text descriptions.
  • GIC requires a model to count either individual objects or semantic groups according to a counting-unit prompt.Examples include counting grapes individually or as bunches, shoes as individuals or pairs, and plates as individuals or stacks.
  • BunchCount provides 1,330 images, 89,254 individual annotations, and 11,065 group annotations with paired labels and explicit containment relations.The relations cover structural arrangement, physical connection, fixed cardinality, and spatial aggregation.
  • Existing counters’ individual-counting capability does not reliably transfer to semantic groups, and naive joint fine-tuning creates a semantic granularity conflict.Improving group counting may degrade individual counting because groups are relational units rather than fixed-appearance objects.
  • GICount uses counting-unit prompts and group-individual containment relations to regularize cross-granularity representations.The framework is designed to improve group counting while preserving individual counting.

Related Work

Prior counting benchmarks broaden target selection through categories, localization, attributes, and context, but continue to treat selected instances as the basic counting units. Related group-perception work studies group structure or constituent-instance regions without making semantic-unit selection the counting target.

  • Early counting benchmarks focused on domain-specific categories, while FSC-147 established class-agnostic counting from target categories.
  • Later benchmarks refine which instances to count using localization, multiple categories, referring expressions, attributes, contextual cues, and distractors.
  • These benchmarks improve target-query semantics but still treat each selected instance as the counting unit.
  • Group-level vision studies aggregate people or localize clusters of same-category objects, while relational models represent relations among object queries.

Group–Individual Object Counting

Group–Individual Object Counting defines counting as selecting either individual instances or semantic groups formed through interpretable relations. The same image and category can therefore yield different valid counts under different prompts.

  • GIC extends object counting by defining semantic groups as higher-level units formed through connection, aggregation, arrangement, or fixed cardinality.
  • BunchCount represents group structures including structural arrangement, physical connection, fixed cardinality, and spatial aggregation.
  • Each group is associated with constituent instances through a containment relation, while some individuals may remain ungrouped.
  • A prompt specifies both the target category and counting unit, which may be individual instances or semantic groups.
  • The same category in one image can have different valid counts, such as individual grapes versus bunches of grapes.

BunchCount Dataset

BunchCount is a real-world benchmark designed for paired individual- and group-level counting in the same images. Its annotations and selection protocol capture diverse, visually identifiable group structures and support explicit group-individual relations.

  • BunchCount provides paired individual- and group-level annotations within the same image.
  • The dataset combines images from established benchmarks and the web, re-annotated under a unified protocol for four group structures.It retains scenes with identifiable instances, clear group boundaries, and reliable group relations.
  • Individual objects are labeled by center points and semantic groups by bounding boxes.
  • Annotation used six annotators over approximately 183 hours, followed by three correction and consistency-checking rounds.
  • BunchCount contains 1,330 images, 89,254 individual point annotations, and 11,065 group bounding-box annotations.Images contain 4–941 individuals and 1–253 groups, averaging 67 individuals and 8 groups.

GICount Framework

GICount combines counting-unit prompts with group–individual relation modeling in a detection-based framework, training on paired individual- and group-level samples. It uses these relations to improve group counting while preserving individual counting ability.

  • GICount is a counting-unit guided relational counting framework built around a counting-unit prompt and group–individual relation modeling.
  • The detection-based architecture localizes and enumerates the semantic unit specified by a textual prompt, rather than regressing a global count.
  • Counting-unit prompts distinguish category semantics from group-structure semantics, representing either individual objects or semantic groups as counting units.
  • GICount obtains counts by enumerating localized predictions, improving group-level counting while preserving individual-level counting capability.
  • Relation modeling aggregates positive and negative individual-query relations and uses metric learning to align group-structure semantics with valid visual organizations.
  • The joint objective combines individual and group detection losses with a relation loss, while paired samples from the same image enable relation-loss computation during training.

Experiments

Experiments show that conventional counting models trade off individual- and group-level performance, while GICount improves group counting and preserves strong individual counting. Ablations further identify containment-based relation modeling and explicit counting-unit prompts as important for avoiding cross-granularity interference.

  • Baseline comparisons: 13.54 individual-level MAE is achieved by CountGD, whereas GroundingREC obtains the strongest baseline group-level result with 3.59 MAE.The corresponding errors are 9.21 group-level MAE for CountGD and 23.17 individual-level MAE for GroundingREC.
  • Baseline comparisons: Conventional instance-counting performance does not reliably predict semantic-group counting ability, as shown by the reversal between CountGD and GroundingREC.CountGD is strongest on individual-level MAE, while GroundingREC is strongest on group-level MAE.
  • GICount results: GICount reaches 1.85 group-level MAE and 5.49 RMSE, reducing GroundingREC's errors by 48.5% and 48.2%.It also obtains 12.20 individual-level MAE and 28.89 RMSE.
  • Ablation study: Adding relation modeling reduces group-level MAE from 9.21 to 1.93, showing that containment-based supervision is crucial for semantic-group counting.Using the full group prompt as the relation anchor instead increases individual-level MAE from 13.54 to 18.36.
  • Counting-units analysis: Individual-only and group-only training specialize by unit, with cross-unit MAEs of 61.73 and 73.95, while direct mixing raises group MAE from 1.38 to 9.21.These results establish counting unit as an essential conditioning factor and expose semantic granularity conflict in joint training.

Conclusions

The paper introduces GIC, BunchCount, and GICount to count either individual objects or semantic groups using the specified counting unit. It establishes a benchmark and baseline for semantic unit-aware counting in practical scenarios such as retail inventory, agriculture, warehousing, and logistics.

  • Contributions: GIC requires models to count either individual objects or the semantic groups they form according to the specified counting unit.This extends visual counting beyond individual instances.
  • Contributions: BunchCount provides paired individual-group annotations and explicit containment relations for the new counting task.The benchmark supports evaluating both counting levels within the same images.
  • Practical scope: The work provides a foundation for semantic unit-aware counting in retail inventory, agriculture, warehousing, and logistics.These are the practical scenarios named in the conclusion.

Semantic Groups and Annotation Scope

The paper treats groups as semantic and perceptual units formed by multiple same-category objects through cues such as arrangement, connection, cardinality, aggregation, or containment. BunchCount uses conservative selection criteria so constituent objects and group membership remain reasonably identifiable and consistent.

  • Semantic group definition: Semantic groups are coherent countable wholes formed through regular arrangement, physical connection, fixed cardinality, spatial aggregation, or shared containment.The paper characterizes groups primarily as semantic and perceptual units rather than products of one geometric rule.
  • Semantic group definition: Group perception can coincide with part-whole organization when repeated objects jointly form a familiar whole.The passage gives beads as an example of this organization.
  • Annotation scope: BunchCount retains scenes where constituent objects are individually identifiable, group boundaries are relatively clear, and membership can be determined consistently.Visible containers, physical connections, regular organization, or separation between neighboring groups can support these boundaries.

Taxonomy of Group Structures

BunchCount organizes semantic groups by the perceptual cues that make multiple individual objects a coherent counting unit, while evaluating transfer across object categories and counting prompts.

  • Group-structure taxonomy: BunchCount defines four group-structure types: structural arrangement, physical connection, fixed cardinality, and spatial aggregation.A group may exhibit multiple cues, but its assigned type reflects the cue most directly supporting its interpretation as coherent.
  • Group-structure taxonomy: Physical-connection groups retain a stable identity through links such as tying, connecting, or joining, despite flexible shapes.Examples include bunches, bundles, strings, and bracelets.
  • Group-structure taxonomy: Fixed-cardinality groups are defined by a conventional or functional number of constituents, such as pairs of shoes.The prescribed cardinality determines how objects are combined across different spatial configurations.
  • Annotation criteria: The benchmark retains only images where individuals and group memberships can be identified clearly, excluding severe occlusion, ambiguous boundaries, and unclear assignments.Groups without enclosing boundaries generally require at least two visible individuals, with exceptions for independently supported single-visible-constituent groups.
  • Evaluation organization: Group-structure descriptions may recur across category-disjoint splits, testing whether learned structures transfer to unseen object categories.The split contains 63 training, 9 validation, and 14 test categories, with 869, 151, and 310 images respectively.
  • Evaluation organization: GICount evaluates both individual and group prompts, with qualitative results reporting predicted response maps for each counting unit and structure-specific errors.Fixed cardinality is the most difficult subset for both CountGD and GICount, while GICount has lower errors on physical connection and structural arrangement.
Loading 2609.04716v1…