Targeted adversarial discriminative domain adaptation

Domain adaptation is a technology enabling aided target recognition and other algorithms for environments targets with data or labeled that scarce. Recent advances in unsupervised domain have demonstrated excellent performance but only when the shift relatively small. We proposed targeted adversarial discriminative (T-ADDA), semi-supervised method extends ADDA framework. By providing at least one image per class, used as cue to guide adaption, T-ADDA significantly boosts of applicable challenging scenario which sets source domains are not same. The efficacy by cross-domain, cross-sensor, cross-target experiments using common digits datasets several aerial datasets. Results demonstrate an average increase 15% improvement over just few images adapting small afforded 60% large shifts.