Previous work has shown that feature maps of deep convolutional neural networks (CNNs) can be interpreted as feature representation of an image. Image features aggregated from these feature maps have achieved steady progress in terms of performances on visual instance retrieval tasks in recent years. The key to the success of such methods is feature representation. In this paper, we study how to represent an image using discriminative features. We demonstrate first that image size is an important factor which affects the performance of instance retrieval but has not been thoroughly discussed in previous work. Based on experimental evaluations, we propose a multi-scale fully convolutional (MFC) approach to encode the image efficiently and effectively. The proposed method is simple to implement, which does not employ sophisticated post-processing techniques such as query expansion, yet shows promising results on four public datasets.