Bipartite Mode Matching Algorithm (BMM)

• BMM is a data-centric algorithm that aligns semantic clusters between target and server domains through minimum-cost bipartite matching.
• It leverages hierarchical clustering and balanced k-means to create mode-balanced training sets that closely mimic target data distributions.
• By employing the Hungarian algorithm for optimization, BMM enhances performance in cross-domain tasks like re-identification and vehicle detection.

The Bipartite Mode Matching Algorithm (BMM) is a data-centric approach that formulates the search for an optimal training set in transfer learning and unsupervised domain adaptation (UDA) as a minimum-cost one-to-one mode alignment problem between semantic clusters (modes) in a target domain and a hierarchical data server. BMM leverages feature clustering, hierarchical representations, and combinatorial optimization (specifically, minimum-weight bipartite matching) to systematically reduce domain gap and enhance downstream task performance. By operating at the distributional mode level—rather than on individual samples—it produces compact, mode-balanced training sets that more faithfully replicate the data distribution of the intended target domain (Yao et al., 14 Jan 2026).

1. Formal Problem Setting and Definitions

Consider a scenario where an unlabeled target dataset $\mathcal{D}_T$ is available, but labeling is infeasible, and a large auxiliary server $\mathcal{S}$ comprising labeled images exists. BMM aims to select a subset $\mathcal{D}_S\subseteq\mathcal{S}$ such that the distribution of $\mathcal{D}_S$ closely approximates that of $\mathcal{D}_T$, thus minimizing the domain gap and facilitating higher accuracy in vision tasks.

Modes are defined as semantic clusters within a feature space, typically extracted via a pre-trained feature extractor $F(\cdot)$ (e.g., ResNet, Inception), mapping each image $x$ to a descriptor $f = F(x) \in \mathbb{R}^d$. The target set $\mathcal{D}_T$ is clustered into $L$ groups $\{\mathbf{T}^i\}_{i=1}^L$, each with mean $\mu^t_i$ and covariance $\Sigma^t_i$, while server data $\mathcal{S}$ is clustered hierarchically into $H$ modes $\{\mathbf{S}^h\}_{h=1}^H$ (Yao et al., 14 Jan 2026).

2. Mode Extraction via Hierarchical Clustering

Server-side, data are first clustered using balanced $k$-means into $J$ clusters (leaves of the hierarchy). An agglomerative merging process forms a binary tree, yielding intermediate clusters representing modes at multiple resolutions, with $H=2J-1$ for a full binary tree. Each mode is characterized by its empirical mean and covariance in feature space.

For targets, flat balanced $k$-means clustering into $L$ groups is used, capturing semantic densities representative of underlying data structure. Hierarchical server organization offers greater flexibility than flat partitions, providing modes of variable size and granularity for alignment with potentially scarce or rare target modes (Yao et al., 14 Jan 2026).

3. Bipartite Matching Formulation and Assignment Algorithm

BMM casts the problem as a bipartite graph assignment: let $X=\{\mathbf{S}^h\}$ (server modes), $Y=\{\mathbf{T}^i\}$ (target modes), $|X|=H$, $|Y|=L$. The cost to align target mode $i$ with server mode $h$ is measured by the Fréchet Inception Distance (FID):

$$\mathrm{FID}(M_t^i, M_s^h) = \|\mu^t_i-\mu^s_h\|_2^2 + \operatorname{Tr}\left(\Sigma^t_i + \Sigma^s_h - 2(\Sigma^t_i\Sigma^s_h)^{1/2}\right)$$

Define binary assignment variables $X_{ih}$. The optimization is:

$$\min_{X_{ih}\in\{0,1\}} \sum_{i=1}^{L}\sum_{h=1}^{H} X_{ih}\, \mathrm{FID}(M_t^i, M_s^h)$$

subject to

$$\sum_{h=1}^H X_{ih} = 1 \quad\forall i\in[L]; \quad \sum_{i=1}^L X_{ih} \leq 1 \quad\forall h\in[H]$$

i.e., each target mode is matched to exactly one server mode, and no server mode is assigned to more than one target mode. The Hungarian algorithm efficiently solves the assignment problem in $O(\max\{H,L\}^3)$ time (Yao et al., 14 Jan 2026).

4. Algorithmic Workflow and Computational Complexity

The procedural steps of BMM are as follows:

1. Feature Extraction: Compute feature descriptors for all images in both $\mathcal{S}$ and $\mathcal{D}_T$.
2. Server Clustering: Balanced $k$-means clusters $\mathcal{S}$ into $J$ clusters; agglomerative clustering forms a hierarchy with $H=2J-1$ modes.
3. Target Clustering: Flat clustering $\mathcal{D}_T$ into $L$ modes.
4. Cost Matrix Computation: For each $(\mathbf{T}^i, \mathbf{S}^h)$, compute FID, leading to an $H\times L$ cost matrix.
5. Bipartite Matching: Solve the assignment via the Hungarian algorithm to obtain the optimal set of matches.
6. Training Set Assembly: Aggregate the assigned server modes $\mathcal{D}_S = \bigcup_{i=1}^L M_s^{h_i}$ (with $h_i$ the assigned server mode for $i$).

Major computational steps are:

• Feature extraction: $O(m_s + m_t)$.
• Balanced $k$-means: $O(m_s J)$ per iteration.
• Agglomerative clustering: $O(J^2)$ with efficient priority queues.
• Final assignment (Hungarian): $O(\max\{H, L\}^3)$. Practical efficiency is achieved by limiting $J$ to 128–256 ($H\sim 256$–512) and precomputing statistics (Yao et al., 14 Jan 2026).

5. Empirical Evaluation and Impact

Experimental results on person/vehicle re-identification (AlicePerson, Market, AliceVehicle, VeRi) and vehicle detection (ExDark, Region100) demonstrate that BMM consistently reduces domain gap (as measured by FID) and yields higher accuracy (Rank-1, mAP) than baseline random selection, standard data pruning, and alternative search methods (e.g., SnP):

Dataset/task	FID (BMM)	Rank-1 (%)	mAP (%)
Person re-ID (Market, 5%)	51.93	49.28	26.08
Vehicle det. (ExDark, 5%)	56.34	–	34.83
Vehicle det. (Region100)	140.07	–	23.08

Hyperparameter analysis reveals stable performance once $J \gtrsim 128$ and $L \gtrsim 10$; hierarchical clustering mitigates the tuning sensitivity seen in flat approaches. The approach is orthogonal to model-centric UDA (e.g., mutual mean-teaching, adaptive teacher). Combining BMM with such methods yields additional performance gains (e.g., Market mAP increases from $\sim$36.05% to $\sim$78.95% when incorporating MMT) (Yao et al., 14 Jan 2026).

6. Relationship to Classical Bipartite Matching and $b$-Matching

BMM's core optimization is rooted in classical bipartite matching, formulated as a minimum-weight assignment on a bipartite cost graph. The broader problem class includes $b$-matching, where each vertex has capacity $b(v)$ and the goal is to maximize total matched weight under degree constraints $1 \leq \deg_M(v)\leq b(v)$. Efficient algorithms for $b$-matching in bipartite graphs (e.g., reduction to perfect matching via vertex replication and Hungarian search, $O(n^3)$ time) exist for richer assignment settings (Rajabi-Alni et al., 2014). While typical BMM selects a perfect matching ($b(v)=1$), the matching methodology aligns with this family, and generalizations could leverage capacity constraints if multi-modal assignment or coverage is desirable.

7. Significance and Application Scope

BMM exemplifies a data-centric paradigm, focusing on optimal construction of matched training sets rather than on iterative adaptation of model weights or pseudo-labels. This allows systematic minimization of domain discrepancy at the cluster/mode level, yielding immediate improvements in generalization for cross-domain recognition tasks (re-ID, detection). Hierarchical server organization enables multi-scale matching, benefiting scenarios with rare or imbalanced modes.

A plausible implication is that BMM can underpin scalable, domain-agnostic transfer with minimal human labeling effort in both vision and potentially other modalities where hierarchical data servers and meaningful distributional features are available. Its orthogonality to model-centric UDA suggests compatibility and composability within larger machine learning pipelines for unsupervised adaptation and transfer (Yao et al., 14 Jan 2026).