EE/CNS 148, CS274c:
Project Suggestions
- Multiresolution Scanning: Imagine a scenario in which one wishes to scan a large statue with millimeter accuracy. No known scanner can do this in one sweep. Instead one would have to acquire individual patches and then stitch these patches together in a global alignment stage. Currently known alignment algorithms cannot deal with this situation. The anology to the numerical optimization problem is that of attempting to build a rigid structure out of lots of stiff tiles glued together with rubber. The entire structure would be rather wobbly... One approach to address this challenge would be to use multiresolution. First acquire a coarse mesh which covers the entire length of the statue (or whatever other object), then acquire high resolution patches and align them to the common low res grid. The potential in this project as as wide as the sky. You can start with a view of the earth and then continually zoom in more until you resolve the wood grain on some window frame...
- Global Alignment Taking into Account Distortion: All current global alignment approaches assume that each partial scan does not contain any distortion beyond scaling. This is rarely the case in practice. Most acquired meshes have some global distortion due to unmodeled effects during the acquisition stage. Instead of trying to model these effects during acquisition one could attempt to build a global alignment tool this allows not just euclidian motions but a somewhat larger class of transformations to be applied to each scan to bring them all into alignment. Can you build such an algorithm?
- Volumetric Calibration: You'll notice that it is quite hard to get globally undistorted point clouds when using only simple camera optics models. One can try to build better models or just acknowledge that no idealized model is going to do a good job and actually calibrate the entire apparatus over a volume figuring out whatever non-linear distortion needs to be compensated for. We have a MicroScribe precision arm which could be used to establish ground truth. How precise can you acquire a model?
- No Structured Light Allowed: What if you can't use structured light to acquire geometry? How much geometry can you infer from a sequence of photographs? Quite a bit. Steven Seitz has suggested the use of voxel coloring to build a voxelated model of an object together with its color appearance by using just a number of photographs with known pose. You could use the microscribe arm with a camera mounted on it to take images with known pose and work from there. Alternatively you could put enough markers into some environment to take images with an uncalibrated camera and recove pose from the markers before doing voxel coloring.
- Exotic Acquisition Techniques: Instead of using a structured light source to infer geometry from active light there are many other ways you can take advantage of the environment to achieve a similar effect. For example, you can use the sun as a light source and a stick to cast a shadow line. Observing its distortion you can triangular just as in the structured light case. What other ideas can you come up with? Imagine using a mirror to reflect sunlight as a thin strip onto the side of a building to scan an entire building that way. Or, take a white light source and use a prism to "colorcode" each of an infinite of stripes by frequency. Or, if you are major laser hacker, build a time of flight laser range scanner... Etc. etc. etc.
- Model Based Acquisition: Consider a class of objects which share certain geometric traits. For example, buildings are generally boxes with possibly triangular prisms for roofs. In the details that's of course not correct, but it is a pretty good start. Using such a model assumption it is possible to build approximate models of buildings for example using just photographs. The rest is done with perspective correct texture mapping. Paul Debevec has demonstrated such a system with great success for architecture. Use these ideas to build models of a significant part of campus. Alternatively think of another interesting class of objects and build a similar system to model those from photographs.