A viable framework for semi-supervised learning on realistic dataset

Semi-supervised Fine-Grained Recognition is a challenging task due to the difficulty of data imbalance, high inter-class similarity and domain mismatch. Recently, this field has witnessed giant leap and many methods have gained great performance. We discover that these existing Semi-supervised Learning (SSL) methods achieve satisfactory performance owe to the exploration of unlabeled data. However, on the realistic large-scale datasets, due to the abovementioned challenges, the improvement of the quality of pseudo-labels requires further research. In this work, we propose Bilateral-Branch Self-Training Framework (BiSTF), a simple yet effective framework to improve existing semi-supervised learning methods on class-imbalanced and domain-shifted fine-grained data. By adjusting stochastic epoch update frequency, BiSTF iteratively retrains a baseline SSL model with a labeled set expanded by selectively adding pseudo-labeled samples from an unlabeled set, where the distribution of pseudo-labeled samples is the same as the labeled data. We show that BiSTF outperforms the existing state-of-the-art SSL algorithm on Semi-iNat dataset. Our code is available at https://github.com/HowieChangehn/BiSTF.