Robust enhancement of depth images from depth sensors

Depth map artefacts reduction: a review

“… Result of [71 ] compared to [69 ], [52 ], and [57 ] for filling hole pixels and filtering the Kinect‐like depth map Baby where (b) is ground truth (reproduced from [71 ]) (a) Simulated kinect depth map baby, (b) Ground truth, (c) DE‐CNN [83], (d) LRM, (e) MSJF, (f) DRECNN [71]…”

“…However, this method is not applicable for real‐time depth processing because it utilises time domain for depth enhancement by taking multiple frame sequences, making it an offline method. An another approach in temporal domain known as 1D least median of squares regression (LMS) is proposed for depth map enhancement [81–83]. This method [83] suggests that the problem of depth pixels flickering through time can be resolved by using a sequence of depth frames to look into invalid depth values and consider them as outliers.…”

“…An another approach in temporal domain known as 1D least median of squares regression (LMS) is proposed for depth map enhancement [81–83]. This method [83] suggests that the problem of depth pixels flickering through time can be resolved by using a sequence of depth frames to look into invalid depth values and consider them as outliers. Then, under some constraints, these invalid pixels are replaced by stable valid depth pixels.…”

A comparative review of plausible hole filling strategies in the context of scene depth image completion

“…Xu et al [124] uses the temporal sequence and motion to create a moving body detection strategy for occlusion filling. More recently, an approach is presented in [151] which uses a sequence of frames to locate outliers with respect to depth consistency within the frame, and utilizes an improved and more efficient regression technique using least median of squares (LMedS) [152] [85] 3.7382 0.0245 51.56e3 ms FMM [41] 1.0117 0.0365 4.31e3 ms DEF [37] 0.6188 0.0030 8.25e5 ms EBI [36] 0.6541 0.0062 9.68e5 ms FBI [75] 0.6944 0.0058 3.84e6 ms DC [113] 0.4869 0.0016 99.09 ms…”

“…• more computationally intensive [8], [7], [91], [80], [83], [116], [123], [113] Depth Image Only • no dependence on extra inputs • less information for processing [104], [107], [108], [135], [136], [109], [9], [110] • more efficient processing • lower quality outputs [86], [11], [117], [118], [12], [90], [151], [41] Table 3: Examples of filling approaches categorized according to the type of images required as their input.…”