Informative sequential selection of variable-sized patches for image retrieval

To quickly and efficiently analyze a large-scale environment by the camera with limited field-of-view, intelligent systems should sequentially select the optimal field-of-view to observe an important and informative patch of area. Especially in the image retrieval task, small observations should be sequentially selected to increase the performance of image retrieval and the updated performance can be used to select the next best view again in a cyclic process. In this paper, we have investigated the role of selected image patches, which can be either overlapped or non-overlapped with previous observations, in this cyclic process. To evaluate the different patch selection strategies, the adaptive observation selection method is also described as follows: (1) robots select adaptive observations sequentially based on its prior knowledge from the training dataset. (2) After each selection, the prior knowledge will be updated by discarding the target-irrelevant data for the next observation selection. During this process, we have shown that an informative patch, even though a part of selected patch is already observed at previous steps, can enhance the retrieval accuracy and it has better performance than an independent observation method.