Single Positive Multi Label

Single Positive Multi-label Learning (SPML) addresses the challenge of training multi-label classifiers using datasets where each image is annotated with only one positive label, significantly reducing annotation costs compared to fully labeled datasets. Current research focuses on developing robust loss functions that mitigate the impact of missing labels, leveraging vision-language models and class priors to generate more accurate pseudo-labels, and employing techniques like contrastive learning and bootstrapping to iteratively refine model predictions. SPML's significance lies in its potential to enable the training of effective multi-label models on large-scale datasets where acquiring complete annotations is impractical, thereby advancing applications in computer vision and other fields requiring multi-label classification.