Segment Based Depth Extraction Approach for Monocular Image with Linear Perspective

In this paper, a segment-guided depth extraction approach is proposed for monocular image with linear perspective. Firstly, foreground depth is learned from a RGBD database with segment-based calibration to adjust the initial coarse depth, and background depth is estimated from linear perspective by vanishing cues. Then, the foreground depth and background one are linearly combined with a statistically optimal balance factor to obtain a holistic fused depth map. Lastly, bilateral filter is exploited to suppress the depth disturbance with edge-preserving. Experiments demonstrate that the proposed technique can produce accurate and dense depths with distinct object boundaries and correct relation among the object positions for a single image.