Parameter Estimation via Deep Learning for Camera Localization

2019 2nd International Conference on Advanced Materials, Intelligent Manufacturing and Automation

Chong, Mina^1 ; Li, Qiming^1 ; Li, Jun^1

Quanzhou Institute of Equipment Manufacturing, Chinese Academy of Science, Quanzhou, China^1

关键词: Camera localization;    Convergence speed;    Convolution neural network;    Illumination changes;    Motion blur;    Network features;    Normalized method;    RGB images;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/569/5/052101/pdf
DOI  :  10.1088/1757-899X/569/5/052101

来源: IOP

【 摘 要 】

This paper proposes a method based on deep learning to estimate parameters for camera localization. The parameters are 6-DOF camera pose and regressed from a single RGB image by an end to end convolution neural network. The proposed network makes use of batch normalized method, which relieves the disappearance of gradient, leading to drastic improvements in convergence speed. In addition, inspired by decomposing convolutions, the network breaks large convolutions into more small convolutions to cascade, which reduces the computational complexity, and enhances network feature representations owing to adding a layer of convolution. This CNN-based method is able to learn effective feature for camera localization in state of motion blur and illumination changes, while the traditional SHIF-based methods fail. Experimental results on both indoor and outdoor public datasets show the improved network achieves an increase in accuracy, and outperforms with the compared methods.