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

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

cs.CVSep 4, 2026

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

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

Organizations: School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China · Shanghai Artificial Intelligence Laboratory, China · Department of Computer Science, University of Liverpool, Liverpool, United Kingdom

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.

Figures & tables

Panel of Figure 1 — Figure 1: Illustration of group-individual object counting. (a) Group-individual object counting. (b) Counting model performance comparisons on the BunchCount.

Figure 2: Representative samples from BunchCount, covering (a) structural arrangement, (b) physical connection, (c) fixed cardinality, and (d) spatial aggregation. Dots and boxes indicate individual and group annotations, respectively.

Panel of Figure 3 — Figure 3: Statistics and comparison of BunchCount. (a) Distribution of group structure types and descriptions. (b) Distributions of the numbers of individual and group annotations per image. (c) Comparison with existing counting datasets.

Figure 4: Overview of the GICount framework. A shared detection backbone receives a counting-unit prompt and predicts the group or individual counts, while the training-only group-individual relation modeling aggregates matched queries.

Method	Prompt Format	Individual Val		Group Val		Individual Test		Group Test
Method	Prompt Format	MAE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$
Qwen3-VL-8B ( 2025 )	Text	34.25	83.37	1.15	5.30	40.07	86.50	10.67	32.21
LLaVA-NeXT-Mistral-7B ( 2024 )	Text	48.59	92.51	3.21	7.17	55.27*	99.42*	11.85	34.51
InternVL3.5-8B ( 2025 )	Text	38.23	85.11	1.56	5.09	46.56	95.97	9.31	30.47
Qwen3.5-27B ( 2026a )	Text	40.00	91.39	0.87	5.33	38.28	82.26	5.84	19.86
GLM-4.5V ( 2026 )	Text	29.90	75.22	0.56	1.52	44.50	131.16	8.83	27.93

Table 1: Main results on the individual- and group-level validation and test sets of BunchCount. The best and second-best results are shown in bold and underlined , respectively. Results marked with * are computed after excluding samples with anomalous model output, details are provided in Appendix B.2.

Method	Individual Test		Group Test
Method	MAE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$
w/o relation modeling	13.54	38.68	9.21	28.30
w/o counting-unit prompt	18.36	57.39	1.93	6.79
full (Ours)	12.20	28.89	1.85	5.49

Table 2: Effect of relation modeling and counting-unit prompt on GICount framework.

Method	Train Setting	Individual Test		Group Test
Method	Train Setting	MAE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$
CountGD	Ind.-only	10.92	26.01	61.73	100.66
	Grp.-only	73.95	115.76	1.38	3.98
	Joint	13.54	38.68	9.21	28.30
GICount (Ours)	Joint	12.20	28.89	1.85	5.49

Table 3: The comparison results under different training settings on the individual- and group-level test sets.

Figure 5: Qualitative results of CountGD and GICount on BunchCount under paired group- and individual-level prompts.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Additional samples from BunchCount.

Split	Classes	Images	Ind. avg.	Grp. avg.
Train	63	869	58	6
Val	9	151	69	5
Test	14	310	90	16

Appendix

Table 4: Split-wise statistics of BunchCount.

Split	Images	Group Structure
Split	Images	Physical connection	Spatial aggregation	Fixed cardinality	Structural arrangement
Train	869	194	237	98	340
Val	151	47	42	6	56
Test	310	39	171	45	55

Appendix

Table 5: Distribution of group-structure types across the dataset splits.

Method	Parameters	Input	FLOPs per Forward
CounTR	99.69M	$384^{2}$	168.66G
CLIP-Count	166.09M	$384^{2}$	54.41G
GroundingREC	173.10M	$800\times 1333$	1.621T
CountGD	237.31M	$800{\times}1333$	1.368T
GICount (Ours)	239.94M	$800{\times}1333$	1.375T / 1.377T

Appendix

Table 6: Computational cost of representative counting methods under their reference input configurations. The two FLOPs values of GICount correspond to individual- and group-level queries, respectively.

Method	Physical Connection (39)	Spatial Aggregation (171)	Fixed Cardinality (45)	Structural Arrangement (55)
CountGD	1.64 / 4.11	0.64 / 1.98	58.22 / 74.05	1.11 / 1.93
GICount (Ours)	1.18 / 1.95	0.74 / 2.24	8.11 / 13.55	0.62 / 1.18

Appendix

Table 7: Group-level counting results across different group structures on the BunchCount test set. Each entry reports MAE / RMSE, and the number of samples in each subset is shown in parentheses.

Method	Individual Test		Group Test
Method	MAE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$
mean-pooling aggregator	16.02	44.13	2.55	8.62
group-query anchor	21.63	69.05	1.51	4.63
triplet loss only	20.37	64.33	1.63	4.49
positive-alignment loss only	18.80	57.10	1.85	5.41
full (Ours)	12.20	28.89	1.85	5.49

Appendix

Table 8: Additional ablations on the components of relation supervision.

Figure 7: Additional qualitative results of GICount on BunchCount.

Object Counting

Object counting estimates the number and locations of target instances in images or video, from dense crowds to fine-grained categories.

See all 61 papers

Count Anything at Any Granularity
May 11, 2026Chang Liu, Haoning Wu, Weidi XieObject Counting Multimodal Large Language Models
School of Artificial Intelligence, Shanghai Jiao Tong University, China · CMIC, Shanghai Jiao Tong University, China
Count Anything
May 29, 2026Mengqi Lei, Shuokun Cheng, Wei Bao +4Object Counting
1{BNRist, THUIBCS, BLBCI, School of Software}, Tsinghua University · School of Computer Science and Technology, China University of Geosciences, Wuhan · 3State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University
UniCounting: Instance-Aware Proposal Consolidation for Image-Query-Free Multi-Category Counting
Oct 6, 2026Jinshi Liu, Pan Liu, Lei He +2Object Counting Data Association
Shenzhen University · The Hong Kong University of Science and Technology (Guangzhou) · Hunan University of Science and Technology +2

May 11, 2026J/K move · Enter open · S save

Count Anything at Any Granularity

Chang Liu, Haoning Wu, Weidi Xie

School of Artificial Intelligence, Shanghai Jiao Tong University, China · CMIC, Shanghai Jiao Tong University, China

Object Counting Multimodal Large Language Models Fine-Grained Visual Perception Multimodal Understanding

Open paper AskPDF arXiv

Open-world object counting remains brittle: despite rapid advances in vision-language models (VLMs), reliably counting the objects a user intends is far from solved. We argue that a central reason is that counting granularity is left implicit; users may refer to a specific identity, an attribute, an instance type, a category, or an abstract concept, yet most methods treat "what to count" as a single, category-level matching problem. In this work, we redefine open-world counting as multi-grained counting, where visual exemplars specify target appearance and fine-grained text, with optional negative prompts, specifies the intended semantic granularity across five explicit levels. Making granularity explicit, however, exposes a critical data bottleneck: existing counting datasets lack the multi-category scenes, controlled distractors, and instance-level annotations needed to verify fine-grained prompt semantics. To address this, we propose the first fully automatic data-scaling pipeline that integrates controllable 3D synthesis with consistent image editing and VLM-based filtering, and use it to construct KubriCount, the largest and most comprehensively annotated counting dataset to date, supporting both training and multi-grained evaluation. Systematic benchmarking reveals that both multimodal large language models and specialist counting models exhibit severe prompt-following failures under fine-grained distinctions. Motivated by these findings, we train HieraCount, a multi-grained counting model that jointly leverages text and visual exemplars as complementary target specifications. HieraCount substantially improves multi-grained counting accuracy and generalizes robustly to challenging real-world scenarios. The project page is available here: https://verg-avesta.github.io/KubriCount/.

Abstract

Object Counting

Explore similar work

Count Anything at Any Granularity

Count Anything

UniCounting: Instance-Aware Proposal Consolidation for Image-Query-Free Multi-Category Counting