Name: GABRIEL SOARES BAPTISTA
Publication date: 27/02/2026
Examining board:
Name |
Role |
|---|---|
| LUCIANO REBOUÇAS DE OLIVEIRA | Examinador Externo |
| LUIS ANTONIO DE SOUZA JUNIOR | Examinador Interno |
| THIAGO OLIVEIRA DOS SANTOS | Presidente |
Summary: Class-Incremental Unsupervised Domain Adaptation (CI-UDA) requires that a model
progressively incorporate new stages containing unlabeled data from a target domain,
where each stage introduces only a subset of the source-domain classes, while retaining
access to labeled source-domain examples throughout the adaptation process. Existing
approaches often struggle to produce truly domain-invariant representations and depend
heavily on post-hoc feature alignment, which can degrade performance, especially under
class imbalance.
To address these shortcomings, this work introduces the Similarity-Driven Attention Pro-
jections (SDAP) method. SDAP exploits similarities among attention-based projections
to produce representation vectors that are discriminative across classes, yet maintain
inter-domain correspondences. This design removes the need to assume balanced class
increments—a requirement difficult to satisfy when dealing with unlabeled data. Further-
more, we incorporate a triplet loss at the instance level to mitigate domain shift, enhancing
intraclass cohesion and interclass separation in the learned feature space.
Empirical evaluation on three established datasets shows that SDAP surpasses current
state-of-the-art CI-UDA methods, achieving 85.8% (±5.2) on Office-Home, 92.1% (±9.3)
on Office-31, and 94.2% (±3.6) on ImageNet-Caltech, representing gains of 13.1, 7.6, and
0.9 percentage points over PLDCA, respectively. These findings confirm the effectiveness of
the proposed approach and establish a new benchmark for CI-UDA, pointing to promising
avenues for future research in incremental unsupervised adaptation.
