This object is challenging due to strong, high-frequency inter-reflections.
Consequently, the depth map using our code ensemble (shown below) has a lot
of residual errors. We propose techniques for detecting and correcting
these residual errors by acquiring more images. These techniques significantly
improve the reconstruction quality. See comparison at the bottom of the page.
|
Error detection using consistency check
|
|
We use a consistency check among the code ensemble for detecting
errors (similar to the one used for handling general scenes).
For a pixel, if no two codes agree, it is marked as an error pixel,
as shown above (marked in red). Since no extra patterns are projected,
the error detection stage places no overhead in terms of acquisition time.
|
|
Error correcting using selective illumination
|
|
In the subsequent iterations, the scene points which are already
decoded correctly are not illuminated. This is achieved using an
illumination mask. By progressively reducing the number of points
getting illuminated (and hence, inter-reflections), the residual
errors are reduced. This technique is based on the work of Xu
et al., 2009.
|
|
Comparison of results
| Conventional Gray Codes(11 images) |
Modulated phase shifting(162 images) |
Our code ensemble(42 images) |
After error detection and
correction(82 images) |
|
|
|