Oceanic Scene Recognition Using Graph-Of-Words (GoW)

Xinghui Dong, Junyu Dong; Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 1122-1130

We focus on recognition of oceanic scene images. A new image dataset is collected. Although it is intuitive to use this dataset to train a CNN from scratch, the limited size of it prevents us from doing so. Instead, it has been shown that encoding the words learnt from deep convolutional features outperforms the fully-connected features extracted using a pre-trained CNN. However, these word encoders do not use the spatial layout of words. As known, this type of data is key to representation of long-range characteristics. Considering graphs are able to encode the complicated spatial layout of nodes, we propose an image descriptor: GoW, to capture the higher order spatial relationship between words. This descriptor is also fused with three word encoders to exploit richer characteristics. These descriptors produce promising results in oceanic scene recognition. We attribute these results to that GoW encodes both the short- and long-range higher-order spatial relationship between words.