next up previous
Next: Research technique Up: Conclusions Previous: Conclusions

Analysis of the system

When performing an eigenvector analysis, it is essential to normalize the input data to correct for variations which the eigenspace should not model; this is the reason, for example, for the alignment stage, which prevents the eigenvectors from attempting to model rotations of the head. We also found that if the input head from the tracking system was significantly brighter than the head models in the eigenspace, the reconstructions were severely deformed, probably because the reconstruction coefficients were being increased beyond their statistically valid range. The experiments performed with the tracking system while modifying the modular eigenspace masks (Chapter 6) indicate that more work is needed to stabilize the image being used as input to the reconstruction pipeline. For example, illumination correction might help the problem of protruding eigenfeatures; however, the complete solution to this problem would involve the full shape from shading computation described in [4], and would probably reduce the reconstruction step from ``interactive time'' (a few seconds) to an offline computation.

The theory behind the modular eigenspace technique (Chapter 5) and experiments done with various modular eigenspace masks (Chapter 6) indicate that with the current technique for combining the modular eigenspaces' contributions, the eigenspaces should be orthogonal to perform properly, but that such orthogonality leads to instability in the reconstruction (tendency to create hard edges in the reconstruction). The diffusion was an attempt to remedy this, but it appears that this causes too great an overlap among the modular eigenspaces, leading to unacceptably large reconstruction error. It may be possible to trade off accuracy for stability by, for example, only blurring the modular eigenspace masks by a very slight amount. Conversely, it may be possible to choose another (nonlinear) combination algorithm.

All of the cross-validation experiments that were performed indicate that we did not have enough data to create a reasonable statistical model. The visual differences between the heads reconstructed with a single eigenspace and with the modular eigenspaces were negligible. In general, as the data set shrinks, the eigenvectors of the data set look more and more like the input vectors rather than like their modes. We found that our eigenheads looked very much like the original data set, as opposed to deformation modes of the head (see [4] for a good example). At the time of this writing we had only 27 head scans to form our eigenspaces. For comparison, the shape from shading experiments done in [4] used a well-known database of 347 CyberWare-scanned air force pilots.

A large amount of time went into rewriting basic input/output code several times, to conform to FLIRT's file formats and to avoid forcing the existing system to change. In retrospect, the existing system should have been modified at the start of the project to use a standard matrix library, which would have allowed the reconstruction system to be implemented in a fraction of the time it took, allowing more development of the underlying algorithms, and leaving less opportunity for error in matrix-related functions. It is recommended that this course of action be taken rather than continuing further development with the existing code base for the reconstruction system.



next up previous
Next: Research technique Up: Conclusions Previous: Conclusions



Kenneth B Russell
Mon May 5 14:33:03 EDT 1997