RGB-Based No-Reference Depth Map Quality Assessment

Zhuo Zhao,Meng Yang

DOI: https://doi.org/10.1109/bmsb53066.2021.9547077

2021-01-01

Abstract:Depth maps play an important role in computer vision and robotics tasks. However, it is difficult to obtain ground truth (GT) of depth maps by either depth sensors or calculation methods. Therefore, there is a great need for depth quality assessment without GT in relevant tasks. This paper proposes a new method to assess the quality of depth map with the guidance of its associated RGB image, which divides the depth map into two parts: structure area and smooth area. On the one hand, it uses the same-view RGB image as a reference to detect the error pixels in depth map structure area, and calculates the objective index as the quality score of the structure area. On the other hand, by using the feature that the pixel values in the smooth area of the depth map are all the same, a sliding window is set to detect the error pixels in the smooth area affected by noise, and the metric is calculated as the quality score of the smooth area. Finally, the assessment scores of the depth map structure area and the smooth area are weighted and integrated as the quality assessment result of the entire depth map. The proposed method is tested on the Middlebury dataset. It can be clearly seen that compared with the assessment results of the other two existing no-reference depth map quality assessment methods, the assessment results of the method proposed in this paper have the strongest consistency with the objective quality assessment results with ground truth (PBMP), and it shows that this method can objectively assess the quality of depth maps.