Method and system for eliminating unsupervised vehicle re-identification deviation based on synthetic data

Abstract

The invention relates to a method and system for eliminating unsupervised vehicle re-identification deviation based on synthetic data. The method comprises the following steps: S1, converting the synthetic data into a pseudo target sample having a similar style with a target domain through a consistent generative adversarial network SPGAN, and carrying out the pre-training of a model through the pseudo target sample; s2, performing an unsupervised domain adaptive or completely unsupervised task through a pre-trained model; s3, calculating the image direction and camera similarity through a pre-trained direction model and camera model, and further obtaining the final vehicle similarity for testing. According to the method and the system, the vehicle re-identification performance is improved, and the adaptability is high.

Description

technical field [0001] The invention belongs to the technical field of vehicle re-identification, and in particular relates to a method and system for eliminating unsupervised vehicle re-identification deviations based on synthetic data. Background technique [0002] With the continuous development of computer vision and Internet of Things, the concept of smart city is promoted. Among them, vehicles, as an important object in smart city applications, have received extensive attention. Since many surveillance cameras are already installed, vehicle re-identification can utilize these cameras to analyze the traffic scene without replacing them with some special hardware. As a cutting-edge and important research topic, vehicle re-identification refers to the retrieval problem of judging whether the vehicle images captured by different cameras in non-overlapping areas belong to the same vehicle in a traffic monitoring scene within a specific range. [0003] Different from previ...

AI Technical Summary

Problems solved by technology

[0006] Although the above-mentioned several mainstream methods for extracting local area features can achieve good results, it can be observed that most of the current vehicle re-identification methods need to label the vehicle images on the original data set, such as the labeling of vehicle key points, Labeling of the local area of ​​the vehicle, labeling of the vehicle attitude direction and other additional labeling information, so there is a lot of information that needs to be labeled

In the real world, it is difficult to collect a data set containing vehicle pictures from different angles, and the number of pictures can reach hundreds of thousands. If these images must be labeled, it is conceivable that the workload is very huge

Moreover, the model that relies too much on annotations has poor generalization. Once the data is changed, the model will not work normally and is not suitable for practical applications.

At present, there are very few studies on unsupervised vehicle re-identification methods. Compared with supervised learning, unsupervised challenges are even greater, such as inter-domain errors caused by different data sets and various perspectives and different vehicles without labels. In-Domain Error due to Orientation

And most of the current methods focus on unsupervised domain adaptation, and the effect in the completely unsupervised domain is not good.

Images