Semi-supervised learning for Fine-Grained Classification with Adaptive Pseudolabeling

Let's break this down:

• The fine-grained image classification task requires subtle variations between visually similar categories—such as differentiating between species of birds with nearly identical features.

• Due to the scarcity of annotated data, supervised approaches tend to perform poorly. Therefore, we rely on semi-supervised learning, which utilizes a small set of labeled images to guide the classification of a much larger unlabeled dataset. 

• Pseudo-labeling is a common semi-supervised learning approach taken to label larger sets of unlabeled data by selecting a threshold to identify reliable predictions, which are then used to enhance the learning process. 

• The threshold significantly Impacts the model's learning process and the quality of labels available to the model. Therefore, identifying the right threshold becomes the crux of the optimization. 

• We propose a variant of adaptive pseudo-labeling that moves the threshold across iterations based on the model's confidence. 

• We propose using class-specific thresholds as it is suboptimal to rely on one threshold to handle the varying characteristics, representations, and difficulties of different classes.