ICAA: Interaction-Calibrated Attention-Based Asymmetric Feature Decomposition and Recovery for Domain Generalization and Adaptation

Normalization narrows visual domain gaps but moves discriminative cues into its residual. Many normalization-residual decomposition methods for domain generalization (DG) and unsupervised domain adaptation (UDA) rely primarily on channel-wise modeling, while existing spatial-channel variants treat spatial and channel cues independently, and symmetric losses judge recovery only against the normalized feature, possibly less discriminative than the original. We propose Interaction-Calibrated Attention-Based Asymmetric Feature Decomposition and Recovery (ICAA). Its Interaction-Calibrated Attention-Based Feature Decomposition and Recovery (IAFDR) module uses Spatial-Channel Interactive Calibration to jointly calibrate spatial and channel evidence and selectively recover task-relevant information from normalization residuals, and an explicit correspondence between convolutional networks and vision transformers extends it to patch tokens. The Asymmetric Task-relevant Feature Decomposition loss encourages recovered features to be more certain than original ones and their complements less certain than normalized ones. ICAA improves every evaluated host across classification, segmentation, and detection: DAEL by 3.9/5.4 percentage points (pp) to 87.3%/71.5% on PACS/Office-Home (1.4/2.2 pp above the best CNN competitor), M3SDA by 9.8 pp to 52.4% on DomainNet, CSFDA by 3.1/4.1 pp to 55.7%/48.7% mIoU on GTA5/SYNTHIA→Cityscapes, and Faster R-CNN by 4.9 pp to 23.7% mAP on Foggy Cityscapes.