Vehicle re-identification (Re-ID) methods with supervised learning achieve high accuracy, but rely heavily on effective supervised labels, so that it cannot extend them to the unsupervised domain. Since the vehicle presents great changes from different perspectives and the large distribution gap between different datasets. Therefore, how to design a vehicle Re-ID method on label-free datasets and show outstanding performance is a difficult problem. In this paper, we propose an unsupervised vehicle Re-ID framework based on synthetic data. Our proposed framework consists of three steps: (1)we use synthetic data to generate pseudo-target samples similar in style to the target domain and use them for model pre-train; (2)the pre-train model is fine-tuned by the source and target domain to improve the cross-domain generalization of the model; (3)the orientation and the camera similarity are calculated by the pre-train orientation and the camera model of the synthetic data, thus punishing the final similarity. Experiments show that the proposed method outperforms existing stateof-the-art methods on benchmark datasets.
Image semantic segmentation plays an important role in assisted driving systems and motor vehicle auto driving system. Due to the complexity of outdoor scenes and driving scenarios, algorithms that only use texture images have low robustness. In order to improve the performance of semantic segmentation, depth images can be used to assist texture images. In addition, the assisted driving system requires that the algorithm need to achieve real-time performance, but the existing algorithm is limited by the complexity of semantic segmentation, resulting in low operating efficiency. To address the above problems, a cross-scale feature extraction module for efficient RGBD image semantic segmentation is proposed. The cross-scale feature extraction module has the characteristics of small parameter amount, large receptive field, and the ability to merge multi-scale features, which can efficiently extract context features. The proposed model achieves a segmentation accuracy of 69.4% mIoU on the RGBD original resolution image of the outdoor scene dataset Cityscapes, and runs at a speed of up to 120 frames per second. Compared with related algorithms, the model proposed in this paper has obvious advantages in running speed, and has achieved a good balance between performance and efficiency.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.