Performance evaluation of whole-image descriptors in visual loop closure detection

We present the performance evaluation of different whole-image descriptors in visual loop closure detection. A whole-image descriptor here is defined as the one that does not require keypoint detection and is therefore fast to extract. In addition, it can be extremely compact to reduce storage requirement. This type of image descriptors are attracting an increasing amount of interest in appearance-based SLAM or robot localization. Our evaluation is in the context of the previous works that have exploited a whole-image descriptor in the application of visual loop closure detection or robot localization. Several whole-image descriptors in three different categories are compared in our study. Our experiments are conducted on two outdoor datasets and the results show that although all these descriptors can be acceptable, they can provide significantly different performance depending upon the evaluation metrics.