Image Generation Based Intra-class Variance Smoothing for Fine-Grained Visual Classification

Fine-grained visual classification (FGVC) is challenging because of the unsmooth intra-class data distribution caused by the combination of relatively significant intra-class variation and scarce training data. To this end, most works in FGVC focused on explicitly/implicitly enhancing the model representation ability. In this paper, however, we take a different stance – alleviating the unsmooth intra-class data distribution in FGVC datasets via data generation. In particular, we propose the following components for data augmentation: (i) SmoothGAN: an information-theoretic extension to the Generative Adversarial Network (GAN) that can generate high-quality fine-grained images with continuously varying intra-class differences. (ii) Dual-threshold-filtering: the generated data are selected according to both their reality and discriminability via SmoothGAN’s discriminator and a basic FGVC model. Experiments on popular FGVC datasets demonstrate that training with augmented data can significantly boost model performance in the FGVC task. The code is available at https://github.com/PRIS-CV/SmoothGAN .