Granular Ball Guided Stable Latent Domain Discovery for Domain-General Crowd Counting

Fan Chen; Shuyin Xia; Yi Wang

cs.CVMar 25, 2026

Granular Ball Guided Stable Latent Domain Discovery for Domain-General Crowd Counting

Authors: Fan Chen, Shuyin Xia, Yi Wang

Organizations: Chongqing University of Posts and Telecommunications, Chongqing, China · Chongqing Ant Consumer Finance Co., Ltd., Chongqing, China

Abstract

Single-source domain generalization for crowd counting is highly challenging because a single labeled source domain may contain heterogeneous latent domains, while unseen target domains often exhibit severe distribution shifts. A central issue is stable latent domain discovery: directly performing flat clustering on evolving sample-level latent features is easily disturbed by feature noise, outliers, and representation drift, leading to unreliable pseudo-domain assignments and weakened domain-structured learning. To address this problem, we propose a granular ball guided stable latent domain discovery framework for domain-general crowd counting. The proposed method first groups samples into compact local granular balls and then clusters granular ball centers as representatives to infer pseudo-domains, thereby converting direct sample-level clustering into a hierarchical representative-based clustering process. This design produces more stable and semantically consistent pseudo-domain assignments. On top of the discovered latent domains, we develop a two-branch learning framework that improves transferable semantic representations via semantic codebook re-encoding and captures domain-specific appearance variations through a style branch, thereby alleviating semantic--style entanglement under domain shifts. Extensive experiments on ShanghaiTech A/B, UCF_QNRF, and NWPU-Crowd under a strict no-adaptation protocol verify the effectiveness of the proposed method and show strong generalization ability, especially in transfer settings with large domain gaps.

Figures & tables

Figure 1: Illustration of three settings under domain shift: fully supervised learning, domain adaptation, and domain generalization.

Figure 2: Overall framework of the proposed method, consisting of semantic codebook re-encoding, style-branch regularization, and granular ball guided stable latent domain discovery for robust density regression.

Figure 3: Hierarchical division and merging of a weighted granular ball set (WGBS) for pseudo-domain discovery. Starting from the initial ball set $S^{(0)}$ , parent balls are iteratively split by weighted 2-means and accepted only when the child compactness decreases sufficiently. After no further split is accepted, $K$ -means is performed on granular ball centers to obtain pseudo-domains.

Source		SHA				SHB				QNRF
Target		SHB		QNRF		SHA		QNRF		SHA		SHB
Method	Publication	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE
DMCount [ 3 ]	NeurIPS’20	23.1	34.9	134.4	252.1	143.9	239.6	203.0	386.1	73.4	135.1	14.3	27.5
D2CNet [ 38 ]	TIP’21	21.6	34.6	126.8	245.5	164.5	286.4	267.5	486.0	–	–	–	–
SASNet [ 39 ]	AAAI’21	21.3	33.2	211.2	418.6	132.4	225.6	273.5	481.3	73.9	116.4	13.0	22.1
MAN [ 40 ]	CVPR’22	22.1	32.8	138.8	266.3	133.6	255.6	209.4	378.8	67.1	122.1	12.5	22.2

Table 1: Comparison with state-of-the-art no-adaptation methods under different source

\rightarrow

target cross-domain transfer settings (MAE/MSE). Venue denotes the publication venue of each method. Results are grouped by the source training set (SHA, SHB, or QNRF) and evaluated on the remaining unseen datasets. Bold values denote the best results.

Source	SHA		SHB		SHA+SHB
Target	NWPU
Method	MAE	MSE	MAE	MSE	MAE	MSE
DMCount [ 3 ]	146.9	563.8	191.6	747.4	144.6	592.8
SASNet [ 39 ]	158.8	588.0	195.7	716.8	155.3	583.6
MAN [ 40 ]	148.2	586.5	193.6	802.5	147.8	605.3
DGCC [ 18 ]	143.1	567.6	175.0	688.6	139.6	553.6

Table 2: Additional evaluation on NWPU-Crowd under a no-adaptation setting. Models are trained on SHA, SHB, or their union, and directly tested on NWPU.

Figure 4: Qualitative comparison on UCF_QNRF under the no-adaptation setting. From left to right: input image, GT density map, and the predictions of Ours, DGCC, UGSDA. The numbers denote crowd counts obtained by integrating the density maps.

Source	SHA				SHB				UCF_QNRF
Target	SHB		QNRF		SHA		QNRF		SHA		SHB
Method	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE
Random partition	12.8	23.4	119.4	218.4	120.2	210.2	169.2	307.1	68.6	114.6	14.3	27.0
K-means	12.9	24.4	119.8	219.6	122.7	220.2	166.1	295.3	68.0	111.3	13.5	25.3
Ours	11.1	15.7	108.0	197.4	97.5	171.7	160.3	298.8	64.1	103.5	13.0	21.5

Table 3: Ablation on pseudo-domain partitioning (MAE/MSE; lower is better). We fix

K

for a clean module ablation (

K{=}4

on SHA,

K{=}3

on SHB, and

K{=}6

on UCF_QNRF); the choice of

K

is discussed in the next subsection. Best results are highlighted in boldface.

Dataset	Method	$\Delta_{\mathrm{med}}\uparrow$	$\sigma_{\mathrm{med}}\uparrow$
SHA	Random	90.8	33.8
	K-means	270.9	110.2
	Ours	278.0	112.7
SHB	Random	20.7	8.5
	K-means	92.9	41.8
	Ours	110.2	46.3

Table 4: GT-count stratification of pseudo-domains. For each method, we compute the median GT count

m_{d}

for each pseudo-domain and report the range

\Delta_{\mathrm{med}}=\max(m_{d})-\min(m_{d})

and the standard deviation

\sigma_{\mathrm{med}}

across the

K

medians.

Source	SHA				SHB
Target	SHB		QNRF		SHA		QNRF
$K$	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE
3	12.8	24.3	120.9	218.1	97.5	171.7	160.3	298.8
4	11.1	15.7	108.0	197.4	119.3	202.9	172.7	306.4
5	12.6	23.7	118.3	215.6	117.4	200.7	177.6	322.0

Table 5: Effect of the number of pseudo-domains

K

under the single-source domain-generalization setting.

Source	SHA				SHB				QNRF
Target	SHB		QNRF		SHA		QNRF		SHA		SHB
Method	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE
baseline	14.8	27.7	134.5	230.5	132.6	229.7	190.6	305.9	88.2	145.0	18.8	35.6
+SCR	13.8	27.2	123.8	228.6	121.7	203.2	171.4	298.9	69.9	116.7	15.4	28.7
+SBR (Ours)	11.1	15.7	108.0	197.4	97.5	171.7	160.3	298.8	64.1	103.5	13.0	21.5
w/o $\mathcal{L}_{sem}$	13.9	27.0	117.8	214.1	119.3	214.8	172.7	306.4	67.2	111.0	14.1	26.5

Table 6: Ablation study of the core components (SCR and SBR) and the SBR regularizers under the single-source domain-generalization setting (without target-domain adaptation).

Crowd Counting

Crowd counting estimates how many people appear in images or video, supporting occupancy monitoring, safety planning, and crowd management.

See all 23 papers

FOCUS: Bridging Fine-Grained Recognition and Open-World Discovery across Domains
Mar 15, 2026Vaibhav Rathore, Divyam Gupta, Moloud Abdar +2Generalized Category Discovery Domain Generalization
Indian Institute of Technology Bombay, Mumbai, India. · Northeastern University, Boston, USA. · The University of Queensland, Brisbane, Australia.
A General Representation-Based Approach to Multi-Source Domain Adaptation
Apr 26, 2026Ignavier Ng, Yan Li, Zijian Li +3Multi-Source Domain Adaptation Unsupervised Domain Adaptation
Carnegie Mellon University · Mohamed bin Zayed University of Artificial Intelligence
Learning Domain Trajectories with Flow Matching for Gradual Domain Adaptation
Oct 22, 2026Yubo Huang, Yushe Cao, Zixi Wang +4Domain Adaptation Unsupervised Domain Adaptation
Key Laboratory of Transportation Tunnel Engineering, Ministry of Education, Southwest Jiaotong University, Chengdu, China · Tsinghua University, Beijing, China · School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, China +2

Mar 15, 2026J/K move · Enter open · S save

FOCUS: Bridging Fine-Grained Recognition and Open-World Discovery across Domains

Vaibhav Rathore, Divyam Gupta, Moloud Abdar, Subhasis Chaudhuri+1

Indian Institute of Technology Bombay, Mumbai, India. · Northeastern University, Boston, USA. · The University of Queensland, Brisbane, Australia.

Generalized Category Discovery Domain Generalization Fine-Grained Visual Perception

Open paper AskPDF arXiv

We introduce the first unified framework for Fine-Grained Domain-Generalized Generalized Category Discovery (FG-DG-GCD), bringing open-world recognition closer to real-world deployment under domain shift. Unlike conventional GCD, which assumes labeled and unlabeled data come from the same distribution, DG-GCD learns only from labeled source data and must both recognize known classes and discover novel ones in unseen, unlabeled target domains. This problem is especially challenging in fine-grained settings, where subtle inter-class differences and large intra-class variation make domain generalization significantly harder. To support systematic evaluation, we establish the first FG-DG-GCD benchmarks by creating identity-preserving painting and sketch domains for CUB-200-2011, Stanford Cars, and FGVC-Aircraft using controlled diffusion-adapter stylization. On top of this ,we propose FoCUS, a single-stage framework that combines Domain-Consistent Parts Discovery (DCPD) for geometry-stable part reasoning with Uncertainty-Aware Feature Augmentation (UFA) for confidence-calibrated feature regularization through uncertainty-guided perturbations. Extensive experiments show that FoCUS outperforms strong GCD, FG-GCD, and DG-GCD baselines by 3.28%, 9.68%, and 2.07%, respectively, in clustering accuracy on the proposed benchmarks. It also remains competitive on coarse-grained DG-GCD tasks while achieving nearly 3x higher computational efficiency than the current state of the art. ^[Code and datasets will be released upon acceptance.]

Abstract

Crowd Counting

Explore similar work

FOCUS: Bridging Fine-Grained Recognition and Open-World Discovery across Domains

A General Representation-Based Approach to Multi-Source Domain Adaptation

Learning Domain Trajectories with Flow Matching for Gradual Domain Adaptation