Abstract
In this work, we combined both a tree and a NetVLAD (vector of locally aggregated descriptors) in order to design new deep model. In developing this model, we studied the impact of tree hyperparameters and found that branch factors had major effects on the parameter utilization as major indicator and the accuracy of the model. The new architecture presented herein exposes a novel hyperparameter called the tree branch factor, which grants additional control over model complexity and on the maps codependency. Deep Tree Net-Vector provide the flexibility to combine two very famous techniques, namely the tree-based technique and VLAD. The former reduces the number of parameters, whereas the latter provides better feature representation inside the model. This work aimed to demonstrate the integration of the strong image descriptor, VLAD, with the tree module, to gain additional control on model size, rather than obtaining better results than state-of-the-art models, and to enhance image representation inside model layers, which could then be invested towards several tasks such as image classification and retrieval. We performed experiments on the Canadian Institute for Advanced Research-10 dataset and were able to show that the proposed models were superior-in terms of information density and accuracy-to many well-known networks.