PROJECT TITLE :

Exploring Spatial Significance via Hybrid Pyramidal Graph Network for Vehicle Re-Identification

ABSTRACT:

Spatial pooling operations are commonly used in existing methods for vehicle re-identification. These operations are used to aggregate feature maps that have been extracted via off-the-shelf backbone networks, such as the visual geometry group network (VGGNet), the Google network (GoogLeNet), and the residual network (ResNet). They fail to investigate the spatial significance of feature maps, which eventually leads to a decline in the performance of the vehicle re-identification system. In the first place, the purpose of this article is to propose a novel spatial graph network (SGN) with the intention of thoroughly investigating the spatial significance of feature maps. Multiple spatial graphs are stacked using the SGN (SGs). Each SG designates certain elements of the feature map as nodes and employs spatial neighborhood relationships in order to determine which nodes connect to one another. In the process of the SGN's propagation, each node and the nodes that are spatially adjacent to it on an SG are merged into the following SG. On the following iteration of the SG, each aggregated node is given a new weighting based on a learnable parameter in order to determine the significance of its corresponding location. Second, in order to investigate the spatial significance of feature maps at a variety of scales in an all-encompassing manner, a novel pyramidal graph network (PGN) has been developed. The PGN provides a hierarchical structure for the organization of multiple SGNs, giving each SGN the ability to manage feature maps at a different scale. The pyramidal graph network (PGN) is then embedded behind a ResNet-50-based backbone network to create a hybrid pyramidal graph network (HPGN). Extensive testing on three large-scale vehicle databases (i.e., VeRi776, VehicleID, and VeRi-Wild) demonstrates that the proposed HPGN is superior to state-of-the-art approaches to vehicle re-identification in terms of accuracy, parameter cost, and computation cost. These results were obtained by comparing the proposed HPGN to the tested approaches. In addition, the results of the experiments demonstrate that the proposed PGN is applicable to all kinds of backbone networks.