A Multimodal Probability Distribution-based Sampling Method for Omnidirectional Depth Estimation

Omnidirectional depth estimation is crucial in enhancing the stability of 3D environment recognition systems. The conventional approach to depth estimation involves predicting a regressed depth value under the assumption of a unimodal probability distribution. However, this method is inherently inadequate for regions containing multiple objects, such as object boundaries. Hence, we propose a novel sampling method based on a multimodal probability distribution and demonstrate its effectiveness by comparing its performance with that of conventional unimodal methods. Based on the experimental results, the proposed method achieves accurate depth prediction in the boundary region, outperforming existing omnidirectional depth estimation models that rely on dense sampling.