Multi-Granularity Class Prototype Topology Distillation for Class-Incremental Source-Free Unsupervised Domain Adaptation

TL;DR

The paper introduces GROTO, a two-module framework for Class-Incremental Source-Free Unsupervised Domain Adaptation (CI-SFUDA) where target data arrive in sessions without access to source data. The Hybrid Knowledge-Driven Positive Class Mining (HKPCM) identifies domain-shared positive classes and generates reliable pseudo-labels using coarse and fine class prototypes, while the Prototype Topology Distillation (PTD) aligns source and target topologies via compactness and separability losses to mitigate forgetting. The approach achieves state-of-the-art results across Office-31-CI, Office-Home-CI, and ImageNet-Caltech-CI datasets and scales to large-scale DomainNet-126-CI, with ablations showing contributions from both modules and their losses. The work provides a practical, privacy-preserving solution for continual, cross-domain learning under label-space shifts, and includes extensive analyses of complexity, training cost, and visualizations. Overall, GROTO advances CI-SFUDA by robustly transferring source knowledge to incrementally arriving target classes while controlling the drift of old knowledge.

Abstract

This paper explores the Class-Incremental Source-Free Unsupervised Domain Adaptation (CI-SFUDA) problem, where the unlabeled target data come incrementally without access to labeled source instances. This problem poses two challenges, the interference of similar source-class knowledge in target-class representation learning and the shocks of new target knowledge to old ones. To address them, we propose the Multi-Granularity Class Prototype Topology Distillation (GROTO) algorithm, which effectively transfers the source knowledge to the class-incremental target domain. Concretely, we design the multi-granularity class prototype self-organization module and the prototype topology distillation module. First, we mine the positive classes by modeling accumulation distributions. Next, we introduce multi-granularity class prototypes to generate reliable pseudo-labels, and exploit them to promote the positive-class target feature self-organization. Second, the positive-class prototypes are leveraged to construct the topological structures of source and target feature spaces. Then, we perform the topology distillation to continually mitigate the shocks of new target knowledge to old ones. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance on three public datasets. Code is available at https://github.com/dengpeihua/GROTO.

Figures (9)

• Figure 1: An illustration of Class-Incremental Source-Free Unsupervised Domain Adaptation (CI-SFUDA) problem, where the labeled source data contain all classes while unlabeled target data come incrementally without access to source instances, and the previously learned target data are unavailable for later adaptations.
• Figure 2: An overview of the GROTO algorithm, it includes two modules: 1) Multi-granularity class prototype self-organization: we mine the positive classes by modeling the source similarity and target probability accumulation distributions, and then promote the positive-class target features self-organization based on the multi-granularity class prototypes via $\mathcal{L}_{ce}$ and $\mathcal{L}_{con}$. 2) Prototype topology distillation: we compute the positive-class prototypes to distillate the topological structures of source and target feature spaces via $\mathcal{L}_{ptd}$.
• Figure 3: The accuracies on old and new classes of GROTO on Office-Home-CI (R$\rightarrow$C), and the average accuracies of 0 to 9 classes for different methods across sessions on Office-Home-CI (R$\rightarrow$C).
• Figure 4: The target cumulative probability and source average similarity results at the first session on Office-31-CI (A$\rightarrow$D). Only training data from 0 to 9 classes are provided at this session. Based on both distributions, we select classes with indices 0-9 as the positive classes at this session.
• Figure 5: The multi-granularity class prototype visualizations of partial classes on Office-31-CI (A$\rightarrow$D), and target feature distribution of source-only, ProCA-B, LEAD-B and GROTO methods on Office-31-CI (D$\rightarrow$A).