TeD-Loc: Text Distillation for Weakly Supervised Object Localization
The episode discusses 'TeD-Loc: Text Distillation for Weakly Supervised Object Localization,' a paper that addresses how weakly supervised models miss full object spatial extent. Hosts discuss using contrastive learning to transfer knowledge from vision-language models like CLIP into patch embeddings, and the use of QR-based orthogonalization to improve class discriminability. The method allows for localization without needing external classifiers during inference.