Vehicle Re-Identification by Deep Feature Embedding and Approximate Nearest Neighbors
Artur O. R. Franco, Felipe F. Soares, Aloísio Vieira Lira Neto, José Antônio Fernandes de Macêdo, Paulo A. L. Rêgo, Fernando A. C. Gomes, José Gilvan Rodrigues Maia · 2020
Disorganized urban growth has led cities to chaos, which has brought countless challenges for their development in several sectors, such as traffic organization, public safety, and transportation. Vehicular re-identification (ReID) technologies have become increasingly important in this context since these allow to produce insights capable of benefiting many areas. As recently, methods frequently resorted to Deep Learning and Convolutional Neural Networks (CNNs), especially in the design of loss functions capable of improving the learning capacity of CNNs. Couple and triplet loss techniques have gained prominence, but their effectiveness depends on the mining of samples to converge properly. In this paper, we investigate a new simple approach for vehicle ReID by combining sample mining strategy and approximate nearest neighbor (ANN) method to improve retrieval quality. By relying only on relatively low-dimensional deep features, we were able to obtain state-of-the-art performance on the VeRi-776 dataset in terms of mAP, HIT@1, and HIT@5 metrics, but using relatively simple CNN and ANN methods, which are feasible in CPU for real-time scenarios.