Performance Comparison of Methods Based on Image Retrieval and Direct Regression for Egocentric Shopping Cart Localization

This work compares the performances of two families of image-based methods for egocentric shopping cart localization in retail stores, namely, methods based on image retrieval and approaches based on direct pose regression from the input image. Our contribution is two-fold: 1) we benchmark the performances of the considered image-based techniques for camera localization in the context of retail stores; 2) we study the computational time and amount of memory required by the considered techniques. Experimental results show that the methods based on image retrieval and the ones based on direct pose regression can achieve comparable localization accuracy, while, especially when a GPU is available, the latter tend to be much faster and require less memory.