A Realistic Evaluation of Semi-Supervised Learning for Fine-Grained Classification
This work addresses the problem of applying SSL to realistic, imbalanced datasets for fine-grained classification, highlighting limitations of current methods and suggesting new strategies, making it incremental in nature.
The paper evaluates semi-supervised learning (SSL) on realistic fine-grained classification datasets with class imbalance and novel classes, finding that SSL methods improve performance when training from scratch but are outperformed by transfer learning, with distillation-based self-training being most robust in transfer settings.
We evaluate the effectiveness of semi-supervised learning (SSL) on a realistic benchmark where data exhibits considerable class imbalance and contains images from novel classes. Our benchmark consists of two fine-grained classification datasets obtained by sampling classes from the Aves and Fungi taxonomy. We find that recently proposed SSL methods provide significant benefits, and can effectively use out-of-class data to improve performance when deep networks are trained from scratch. Yet their performance pales in comparison to a transfer learning baseline, an alternative approach for learning from a few examples. Furthermore, in the transfer setting, while existing SSL methods provide improvements, the presence of out-of-class is often detrimental. In this setting, standard fine-tuning followed by distillation-based self-training is the most robust. Our work suggests that semi-supervised learning with experts on realistic datasets may require different strategies than those currently prevalent in the literature.