Multi-task Relative Attributes Prediction by Incorporating Local Context and Global Style Information Features

Relative attribute represents the correlation degree of one attribute between an image pair (e.g. one car image has more seat number than the other car image). While appearance highly and directly correlated relative attribute is easy to predict, fine-grained or appearance insensitive relative attribute prediction still remains as a challenging task. To address this challenge, we propose a multi-task trainable deep neural networks by incorporating an object’s both local context and global style information to infer the relative attribute. In particular, we leverage convolutional neural networks (CNNs) to extract feature, followed by a ranking network to score the image pair. In CNNs, we treat features arising from intermediate convolution layers and full connection layers in CNNs as local context and global style information, respectively. Our intuition is that local context corresponds to bottom-to-top localised visual difference and global style information records high-level global subtle difference from a top-to-bottom scope between an image pair. We concatenate them together to escalate overall performance of multi-task relative attribute prediction. Finally, experimental results on 5 publicly available datasets demonstrate that our proposed approach outperforms several other state-of-the-art methods and further achieves comparable results when comparing to very deep networks, like 152-ResNet [19] and inception-v3 [8].