Modeling Kinect Sensor Noise for Improved 3D Reconstruction and Tracking

Elastic Fragments for Dense Scene Reconstruction

“…The two error models we use aim to simulate the factory-calibrated data produced by PrimeSense sensors and idealized data with no low-frequency distortion. To produce the idealized data, we process the perfect synthetic depth images using the quantization model described by Konolige and Mihelich [13] and introduce sensor noise following the model of Nguyen et al [18]. To produce the simulated factorycalibrated data, we add a model of low-frequency distortion estimated on a real PrimeSense sensor using the calibration approach of Teichman et al [31].…”

Optimal 6D Object Pose Estimation with Commodity Depth Sensors

“…The first data set corresponds to flat walls in order to obtain axial errors and the second corresponds to a checkerboard block grid in order to obtain axial errors. As is assumed in most other depth error models ( [72], [80], etc. ), the error from each pseudo-measurement transformed from a depth image is represented as a unique multivariate Gaussian distribution.…”

“…In Figure 3.2(C), the hypersurface fits are plotted on the left for the sensor positioned in front of the wall at a depth of 1600 mm, 2400 mm, and 3200 mm. On the right, the standard depth errors are scatter plotted as a function of radial distance from the center of the focal plane, along with the Menna [72], Nguyen [80], and recalibrated Choo models. The Menna and Nguyen models clearly misrepresent the standard depth errors by assuming a constant, linear model for depth estimates across the focal plane.…”

“…In contrast, Nguyen et al [80] developed empirical models for axial and lateral error, and explored the effects of flat tilted surfaces on depth estimation. In doing so, they added a hyperbolic term to their quadratic axial and linear lateral models by fitting parameters to experimental data.…”

“…It should be noted that due to the nature of empirical models, the estimates of model coefficients are strongly dependent on the measured data sets, and errors from any hidden variables not accounted for will be mixed into the estimated parameter values. In the case of [80], factors such as surface material properties and ambient thermal conditions are implicitly kept constant since they are not modeled. Also, since depth data was collected on surfaces near the center of the optical axis, error due to lens distortion was not considered.…”