Recurrent Omnidirectional Stereo Matching Method Based on Mixture of Laplace Loss

Recurrent Omnidirectional Stereo Matching (RomniStereo) is an effective method to estimate omnidirectional depth. At present, most of the depth estimation methods ignore the distortion of the fisheye images, which has a negative impact on the performance of the algorithm. To deal with distortion at the edges of the fisheye image, a new mixture of Laplace loss function is used in the network, which add the uncertainty estimation of the pixels on the fisheye image in traditional loss function. Uncertainty estimation can effectively distinguish the distortion degree of fisheye images, so as to improve the depth estimation effect under the condition of large distortion. There are three important parameters in loss function are predicted and updated through loops for greater accuracy. Experimental results show that the proposed method generates excellent results in both synthetic and real-world environments, and the distortion at the edges of the fisheye image has been effectively improved.