Spatially-Excited Attention Learning for Fine-Grained Visual Categorization

Learning distinguishable feature embedding plays an important role in fine-grained visual categorization. The existing methods focus on either designing a complex attention mechanism to boost the overall classification performance or proposing a specific training strategy to enhance the learning of the backbone network to achieve a low-cost backbone-only inference. Unlike all of them, an alternative approach called Spatially-Excited Attention Learning (SEAL) is proposed in this paper. The training of SEAL is similar to that of most of the existing methods, but it provides two alternative streams during a network inference: one stream requires higher effort but provides higher performance; the other is a low-cost backbone-only inference with lower but still comparative performance. Note that both the streams are trained at the same time by SEAL. The experiments show that SEAL achieves the state-of-the-art performance under both complex architecture and backbone-only inference conditions.