Cross-Modal Knowledge Transfer Without Task-Relevant Source Data

AbstractCost-effective depth and infrared sensors as alternatives to usual RGB are now a reality, have some advantages over in domains like autonomous navigation remote sensing. As such, building computer vision deep learning systems for data crucial. However, large labeled datasets these modalities still lacking. In such cases, transferring knowledge from neural network trained on well-labeled dataset the source modality (RGB) that works target (depth, infrared, etc.) is of great value. For reasons memory privacy, it may not be possible access data, transfer needs work with only models. We describe an effective solution, SOCKET: SOurce-free Cross-modal KnowledgE Transfer this challenging task one different without task-relevant data. The framework reduces gap using paired task-irrelevant well by matching mean variance features batch-norm statistics present show through extensive experiments our method significantly outperforms existing source-free methods classification tasks which do account gap.KeywordsSource free adaptationCross modal distillationUnsupervised domain adaptation