Multi Stage Quasi-labeled Vehicle Re-Identification Using Branched Convolutional Neural Networks

Abstract

Vehicle re-identification has become an important field of interest in computer vision, especially after the rise of importance of smart cities, and also with the evolution of Convolutional Neural Networks it has seemed possible to take advantage of technology to perform vehicle re-identification, Due to vehicles being an important object in the formation of Smart cities many scholars have applied different techniques to perform vehicle re-identification as it has many applications such as unmanned vehicle parking, vehicle tracking, vehicle identification, and vehicle counting and also arial traffic flow management. Many scholars have tried to solve vehicle re-identification challenges using computer vision conventional methods such as feature extraction and histogram of oriented gradients but with the evolution of CNN which itself tries to learn the representation of the vehicle from images with the labels provided and this method has proved to be very successful instead of its predecessors and has achieved great results in the field of Re identification. Person re-identification as well as vehicle re-identification. Although supervised vehicle re-identification has achieved great results in vehicle re-identification but due to the complication of implementing supervised re-identification methods which are huge data annotation costs and due to their classification nature supervised methods are always in need of finetuning on new data to produce good results, so we have tried to solve this problem with unsupervised learning combining it with the training of CNN’s using the technique of pseudo label generation which are used to learn representation for CNN’s and they are gathered from data using some conventional clustering techniques used for unsupervised learning. Vehicle reidentification has different challenges present in real world scenarios and one of the major problems is vehicle appearance or different views of vehicles in different cameras which can be described as domain discrepancy in different cameras we have tried to solve that problem for vehicle re-identification using unsupervised learning by training a CNN in for intra camera to adjust model feature xii generalization across the camera and then fine-tuning backbone for inter camera training to further make the model backbone robust to view discrepancy present in the real world. We have also infused vehicle contextual information (color, type) information in the model for the model to better learn the representation and learn class discrimination for a single vehicle in different cameras and to converge faster. We First trained the color model in the same manner as we are training the re-id model and then after training, using an unsupervised hierarchical technique for clustering we produced pseudo labels and fed those pseudo labels for the color label to Re-id Network after converting it into the multi-class multilabel network. Idea was to increase feature similarity between same class boundary learning and decrease feature similarity between different class samples.