Abstract:
Fine-grained visual classification is inherently challenging because of its inter-class similarity and intra-class variance. However, by contrasting the images with same/different labels, a human can instinctively notice that the key clues lie in certain objects while other objects are ignorable. Inspired by this, we propose Contrastively-reinforced Attention Convolutional Neural Network (CRA-CNN), which reinforces the attention awareness of deep activations. CRA-CNN mainly contains two parts: the classification stream and attention regularization stream. The former classifies the input image and simultaneously divides the visual information of the input into attention and redundancy. The latter evaluates the attention/redundancy proposal by classifying the attention and contrasting the attention/redundancy of various inputs. The evaluation information is backpropagated and forces the classification stream to improve its awareness of visual attention, which helps classification. Experimental results on CUB-Birds and Stanford Cars show that CRA-CNN distinctly outperforms the baselines and is comparable with state-of-art studies despite its simplicity.
