WACV 2027Algorithms Track COMPUTER VISION RESEARCH

DGA-CAFC · DOMAIN-ROBUST VISUAL RECOGNITION

Attention Drift Regularization with Adaptive Initialization for Domain-Robust Visual Recognition

Shivangi Rai·Vinod K Kurmi

Department of Data Science and Engineering
Indian Institute of Science Education and Research Bhopal, India

▤ Paper Coming Soon⌘ Code Coming SoonExplore the framework ↓
01 / ABSTRACT

Build robust representations beyond source domains.

Source-free domain adaptation (SFDA) adapts a source-trained model to unlabeled target domains without access to the source data. Two important failure modes remain: attention drift, where learned attention shifts toward spurious or background cues under distribution shift, and domain-gap-agnostic initialization, where the same source model is used irrespective of the target-domain shift.

We propose DGA-CAFC, which combines Counterfactual Attention-based Feature Consistency (CAFC) and Domain-Gap Aware Source Model Selection (DGASS). CAFC stabilizes attention-conditioned features using counterfactual-inspired normalized attention perturbations, while DGASS uses prediction entropy on unlabeled target data to select an appropriate source-model variant before adaptation.

CAFC is also applicable as a lightweight regularizer for domain generalization (DG), where target-domain data are entirely unavailable during training.

Source-Free Domain AdaptationDomain GeneralizationAttention Consistency
02 / FRAMEWORK

Two modules. One robust framework.

DGASS selects the initialization before adaptation; CAFC regularizes learned features throughout training.

FIGURE 2FULL DGA-CAFC FRAMEWORK
DGA-CAFC full framework depicting source variants, entropy-guided model selection, teacher-student adaptation, and CAFC loss
Overview of DGA-CAFC: source-model candidates are selected according to target prediction entropy, and attention-conditioned representations are regularized during target adaptation.
03 / METHODOLOGY

Make attention stable. Make initialization adaptive.

MODULE / 01
◎

CAFC

Counterfactual Attention-based Feature Consistency

Builds uniform, random, and reversed normalized attention maps. Aligning features computed from learned and counterfactual attention encourages robustness to attention drift.

Feature stability · SFDA + DG
MODULE / 02
◇

DGASS

Domain-Gap Aware Source Model Selection

Measures uncertainty via target prediction entropy and selects between complementary source-model variants, providing an initialization suited to the estimated domain-gap severity.

Entropy guidance · SFDA
FIGURE 1CAFC ATTENTION CONSISTENCY
Illustration of counterfactual attention-based feature consistency
CAFC encourages feature consistency between learned attention and counterfactual-inspired perturbations, reducing sensitivity to spurious attention shifts.
04 / EXPERIMENTAL RESULTS

Measured across five benchmarks.

Values shown below are taken from the WACV 2027 manuscript. All chart values are mean top-1 classification accuracy (%).

91.65%Office31 · SFDA
79.75%Office-Home · SFDA
87.49%DomainNet40 · SFDA
SOURCE-FREE DOMAIN ADAPTATION

Baseline comparisons

Hover over bars for exact values. Select a benchmark to switch datasets.

Baselines DGA-CAFC (ours)
DOMAIN GENERALIZATION

CAFC as a plug-in regularizer

ResNet-50 (ImageNet-pretrained); average accuracy for each objective.

Original objective With CAFC

SFDA: average over transfer tasks. DG: average over held-out domains; numerical means are shown for the ImageNet-pretrained ResNet-50 setting. The manuscript reports DG standard deviations separately. Charts omit detailed task-level results for clarity.

05 / CONTACT

Get in touch.

For questions about DGA-CAFC, contact the authors.