Maya Skies Internal Production Site

Unified Test B

Written by Administrator

Monday, 13 August 2007

Overview

In this second test, we use a clay model as our test subject. We do this in order to simplify the stereo reconstruction task (in comparison with our first test -- see "Unified Test A") and to provide deeper bas relief for capture. For this test we sculpted a small test subject in clay.

Photometric Image Data

Using one camera in the stereo pair, we acquired 7 images with a remote strobe in roughly compass point positions relative to the test surface. In the proposed hybrid system, it is unclear whether we should continue to use a manually controlled satellite strobe. The advantage of using fixed lighting (as Holly Rushmiere did with her MR16 sources for IBM's Pieta project) is that fiducials should not be required in order to accurately locate the light sources. This should accelerate capture in the field. The disadvantages are: 1. Each strobe source can have a slightly different output profile, which must then be calibrated for. 2. In order to be practical, only 4 or 5 strobes are possible, which may not be enough to achieve results for all pixels in the scene. 3. Having many strobes in outriggers adds to the weight and bulk of the capture system. As shown below, fiducials are inserted in the scene in order to locate the strobe source in the scene.

A rectified photometric source image:

Download the rectified photometric image set: Rectified Photometric Source Images (~16MB)

The following images, below, show all seven input images. The ambient plate is fairly noisy given the 800 ISO value for these photos. In the production system Mark and I are shooting for a 100 ISO setting. Seven input images for photometric stereo processing: six strobe positions + ambient plate:

Download the photometric image set: Unified Test B Images (raw source) (~16MB)

Limitations of the Image Data

Despite shooting at an "ideal" focal distance of ~70mm and highly uniform depth in the scene, some regions of the image are out of focus.
Since the test system we used is hard-wired to use ISO 800, grain is apparent in the images.
Comparatively high sampling rate on ground truth. Total ground truth samples, in fact, are approximately the same the total photometric image samples.

Ground truth Range Data

In addition to the stereo matching and photometric stereo image capture, we used a commercial close-range scanner to create a "ground truth" record. A view of this model data is below. This model has been aligned with the model we obtained via stereo reconstruction.

Three views of the ground truth data: shaded mesh, Z-depth and normal map:

click for a larger view

Stereo Image Capture

All images were captured using the following system:

(2) Pentax K110D 6.1MP Digital SLR Cameras with 18-55mm f/3.5-5.6 Lens
Image size: 3008 x 2000
F4.5, Focal length: 43mm

A structured light pattern is projected into the scene using a customized strobe system.

From left to right: textured low-res mesh, shaded low-res mesh, shaded mid-res mesh, shaded high-res mesh:

Download the low-res & mid-res stereo data (OBJ): Unified Test B Low & Mid Models (~7MB)

The "high res" mesh computed from stereo matching results is shown below, aligned with the ground truth data above.
Aligned stereo data with ground truth:

Download the depth map from stereo correspondence (ASCII, TIFF): Depth Map (~13MB)
Download the aligned stereo data & ground truth models (OBJ): Unified Test B Models (~35MB)

Geometric Reconstruction Results

The goal of our geometric reconstruction is a modeling process that combines the accurate high-frequency normal data provided by the photometric stereo process with the accurate low-frequency data provided by the stereo matching process. Using a range map from the stereo correspondence data and a normal map from the photometric data, a geometric reconstruction of the scene is obtained. Below are images of the initial geometric reconstruction.

Orthographic front view: (left to right) ground truth, geometric reconstruction, stereo reconstruction.

click for a larger view

In the above image, the initial stereo correspondence (above, right) aligns well with the ground truth (above, left). However, there are obvious differences when comparing either the ground truth or stereo model with the initial geometric reconstruction (above, center). This could be a result of problems with the range map provided as one of the inputs to the geometric reconstruction, described above.
Orthographic front view: (left to right) geometric reconstruction v. stereo reconstruction, ground truth v. geometric reconstruction, ground truth v. stereo reconstruction.

click for a larger view

Above, we see the reasonably close alignment possible between the ground truth and stereo correspondence models. The delta between the geometric reconstruction and ground truth (above, center) is much greater, and an ICP alignment between the two models failed to converge.

Comments

Only registered users can write comments!

Last Updated ( Sunday, 15 June 2008 )

< Prev		Next >

[ Back ]