💻 computer science

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

This paper proposes a granular ball guided framework that achieves stable latent domain discovery through hierarchical representative-based clustering and a two-branch learning strategy to disentangle semantic and style features, thereby significantly improving generalization in single-source domain-general crowd counting.

Original authors: Fan Chen, Shuyin Xia, Yi Wang, Xinbo Gao

Published 2026-03-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fan Chen, Shuyin Xia, Yi Wang, Xinbo Gao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a crowd counter. Your job is to look at a photo of a crowd and guess exactly how many people are in it.

In the real world, this is hard. A photo taken at a sunny beach in Brazil looks very different from one taken in a rainy subway station in Tokyo. The lighting, the camera angle, and how packed the people are all change. If you train a computer to count people using only photos from the sunny beach, it will likely fail miserably when you show it the rainy subway.

This is the problem of Domain Generalization: teaching a computer to be smart enough to handle any new environment it hasn't seen before, without needing to re-train it.

Here is how this paper solves that problem, explained simply:

1. The Problem: The "Noisy Classroom"

The researchers realized that even the "training data" (the sunny beach photos) isn't just one big, uniform group. It's actually a mix of many different "micro-environments" (e.g., some photos are close-ups, some are far away, some are crowded, some are sparse).

Old methods tried to group these photos by looking at every single photo individually. But this is like trying to organize a messy classroom by asking every single student to raise their hand and shout out their group. It gets chaotic! A few noisy students (outliers) or a student who is having a bad day (feature drift) can mess up the whole grouping. The computer gets confused and learns the wrong things.

2. The Solution: The "Granular Ball" Strategy

The authors propose a new way to organize the data, which they call Granular Ball Guided Stable Latent Domain Discovery.

The Analogy: The "Class Representative" System
Instead of asking every single student (photo) to vote, the computer first groups students into small, tight-knit clusters called "Granular Balls."

Think of a Granular Ball as a small group of friends sitting at a lunch table. They all look similar.
Instead of listening to 100 individual voices, the computer picks one "Class Representative" (the center of the ball) to speak for the whole group.
Then, it organizes the representatives, not the individuals.

Why is this better?
If one student in the lunch group is shouting nonsense (noise), the representative can still speak calmly and accurately for the whole group. This makes the grouping much more stable and less likely to be thrown off by mistakes. It turns a chaotic flat list into a structured, hierarchical system.

3. The Two-Brain System

Once the computer has organized the photos into these stable groups, it uses a special Two-Branch Learning Framework to learn how to count. Imagine the computer has two different "brains" working together:

Brain A (The Semantic Brain): This brain focuses on the meaning of the crowd. It asks: "How many heads are there?" It uses a special Codebook (like a dictionary of crowd patterns) to make sure it understands the concept of "crowd" no matter where the photo was taken. It ignores the background noise.
Brain B (The Style Brain): This brain focuses on the look of the photo. It asks: "Is it sunny? Is it rainy? Is the camera grainy?" It learns the "style" of the specific group (the Granular Ball) so it doesn't get confused by the lighting.

The Magic Trick:
The system forces these two brains to stay separate. Brain A learns the count, and Brain B learns the style. They are told, "Don't let the style mess up the count!" This prevents the computer from thinking, "Oh, this photo looks like a rainy day, so I must be wrong about the number of people."

4. The Result: A Super-Adaptable Counter

By using this "Class Representative" system to find hidden groups and then teaching the computer to separate "counting" from "style," the model becomes incredibly robust.

The Test: They tested it on famous crowd datasets (like ShanghaiTech and UCF QNRF).
The Outcome: Even when the computer was trained on one type of crowd and tested on a completely different type (e.g., from a sparse street to a massive stadium), it performed better than all previous methods. It didn't get confused by the change in scenery.

Summary

Think of this paper as teaching a computer to be a wise detective rather than a confused tourist.

Old way: The tourist looks at every single person in a crowd and tries to guess the total, getting overwhelmed by the chaos.
New way: The detective groups people into logical clusters, listens to the group leaders, separates the "who" (the count) from the "where" (the style), and gives a confident, accurate answer no matter the environment.

This approach makes crowd counting reliable for real-world applications like traffic management, event safety, and city planning, even in places the computer has never visited before.

1. Problem Statement

The paper addresses Single-Source Domain Generalization (DG) for crowd counting.

The Challenge: Crowd counting models typically fail when deployed on unseen target domains due to significant distribution shifts (e.g., changes in density, viewpoint, illumination, and scene layout).
The Specific Bottleneck: Existing DG methods attempt to discover "latent domains" (pseudo-domains) within a single labeled source dataset to improve robustness. However, current approaches rely on flat clustering of sample-level features. This is highly unstable because evolving feature representations during training are susceptible to noise, outliers, and representation drift, leading to unreliable pseudo-domain assignments.
The Goal: To develop a method that stably discovers latent domain structures within a single source dataset and learns disentangled representations (separating counting semantics from domain-specific styles) to generalize to unseen targets without adaptation.

2. Methodology

The proposed framework, Granular Ball Guided Stable Latent Domain Discovery, consists of three main components:

A. Granular Ball Guided Stable Latent Domain Discovery

Instead of clustering individual samples directly, the method employs Granular Ball Computing (GBC) to create a hierarchical, robust clustering process:

Descriptor Extraction: Domain-sensitive descriptors are extracted from multi-level feature statistics (mean and standard deviation) of the backbone network.
Hierarchical Partitioning:
- Samples are recursively partitioned into compact granular balls (local structural units) using a weighted 2-means algorithm.
- Splitting is controlled by a compactness metric; a split is accepted only if it significantly reduces the weighted deviation of the child balls compared to the parent.
- This transforms the problem from unstable sample-level clustering to a hierarchical representative-based clustering.
Pseudo-Domain Assignment: The centers of the final granular balls serve as stable representatives. Standard K-means is applied to these ball centers (rather than raw samples) to generate $K$ pseudo-domains.
Label Alignment: To ensure consistency across training epochs, the newly discovered pseudo-domain labels are aligned with the previous epoch's labels using Hungarian matching.

B. Two-Branch Learning Framework

Based on the discovered latent domains, the model uses a shared backbone with two specialized branches:

Semantic Branch (Counting Core):
- Focuses on domain-invariant crowd semantics.
- Employs a Semantic Codebook Re-encoding mechanism. Features are projected and re-encoded via a learnable codebook using soft assignment to create a compact, stable representation for density regression.
- Loss ( $L_{sem}$ ): Enforces consistency of semantic descriptors across different pseudo-domains.
Style Branch (Regularization):
- Captures domain-specific appearance variations (style).
- Does not contribute to density prediction but provides regularization signals.
- Loss ( $L_{sty}$ ): Enforces intra-domain style compactness (samples in the same pseudo-domain should have similar style features).
- Loss ( $L_{orth}$ ): Enforces orthogonality between semantic and style features to prevent information leakage (disentanglement).

C. Training Objective

The total loss function combines density regression with the proposed regularizers:
$\mathcal{L} = \mathcal{L}_{den} + \lambda_{sem}\mathcal{L}_{sem} + \lambda_{sty}\mathcal{L}_{sty} + \lambda_{orth}\mathcal{L}_{orth}$
Where $\mathcal{L}_{den}$ is the standard density map regression loss.

3. Key Contributions

Identification of Instability: The authors identify that direct sample-level flat clustering is a fundamental bottleneck in single-source DG due to feature noise and drift.
Granular Ball Guided Discovery: They propose a novel strategy that aggregates samples into granular balls first, then clusters ball centers. This hierarchical approach significantly stabilizes pseudo-domain assignments and is robust to representation drift.
Disentangled Representation Learning: A two-branch framework is developed that uses the discovered latent domains to enforce semantic consistency, style compactness, and semantic-style orthogonality, effectively reducing domain leakage.
Data-Driven Hyperparameter Selection: A simple rule ( $K \approx N^{1/4}$ ) is proposed to automatically determine the optimal number of pseudo-domains based on dataset size, reducing the need for extensive hyperparameter tuning.

4. Experimental Results

The method was evaluated on standard benchmarks: ShanghaiTech (A/B), UCF QNRF, and NWPU-Crowd under a strict no-adaptation protocol.

Performance: The proposed method consistently outperforms strong baselines (including DGCC, DMCount, and adaptation-based methods) across all cross-domain transfer settings.
- Example (SHB $\to$ QNRF): Improved MAE from 179.1 (DGCC) to 165.3.
- Example (QNRF $\to$ SHA): Achieved the best MAE of 64.1, significantly outperforming previous state-of-the-art.
- NWPU-Crowd: Showed consistent improvements in MSE, indicating a reduction in extreme counting errors (outliers) on highly heterogeneous scenes.
Ablation Studies:
- Granular Ball vs. K-means: The granular ball strategy yielded lower MAE/MSE than direct sample-level K-means or random grouping, proving the stability of the hierarchical approach.
- Density Regime Separation: The granular ball method successfully separated pseudo-domains based on crowd density regimes (sparse vs. dense), which is crucial for effective regularization.
- Component Analysis: Removing either the Semantic Codebook Re-encoding or the Structured Branch Regularization (SBR) led to performance drops, confirming the necessity of both semantic enhancement and style disentanglement.

5. Significance

Robustness to Distribution Shift: The method provides a reliable solution for deploying crowd counting models in real-world scenarios where target data is unavailable or privacy-restricted.
Structural Modeling: It introduces a new perspective to domain generalization by shifting from "flat" sample clustering to "hierarchical" structural modeling using granular balls, which is more resilient to the noise inherent in deep learning feature spaces.
Disentanglement: By explicitly separating counting semantics from domain styles using the discovered latent structure, the model learns more transferable representations, addressing a key limitation in current DG literature.
Practicality: The approach requires no target data during training and offers a systematic way to determine the number of latent domains, making it highly applicable to diverse deployment scenarios.