Texas Database

The Texas database is now ready for use (Texas3DFR Database) to get the data. The data must not be used or copied for any usage other than our academic research. This dataset ought to be very suitable for our needs because it is prealigned rigidly, which removes some of the issues encountered so far. It also makes ICP-agnostic comparisons (based on non-rigid recognition alone) easier. Their dataset was used by several US universities (but not many) and it is apparently much higher in terms of its quality. They are still trying to get more groups to use it. We had been waiting for a response regrading access to the new files. It was work in progress. It would be great to see performance attained from pre-aligned data. A lot of the current difficulties are associated with this drawback.

Data preparation works as before on the Texas dataset, having encoded the necessary adjustments and make them more modular (so as to keep the program compatible with GIP datasets and NIST datasets). Currently, the program crashes at GMDS sometimes, but that ought to be resolvable within days.

With caveats, after much work at increased pace, the GMDS-PCA code can now be applied to parts of the data from Texas, as shown in figures $[*]$ , $[*]$ , and $[*]$ . These are not matches between identical people but between different people. The code is not yet in a state good enough for benchmarks.

**Figure:** Standard program settings with which to run the Texas data

**Figure:** An example of GMDS applied to just a vertical slice of the data taken from different individuals

**Figure:** Exploratory work around GMDS applied solely to the nose region of different people (left and right), shown from different angles

After much debugging that includes visualisation, regression and repetition, it appears likely that the bug we have been investigating for weeks is in fact not truly a bug in my code but an initialisation problem associated with GMDS. When bouncing back and forth between the two surfaces trying to match one to another, GMDS sometimes appears to fall flat on its face. When it gets things right, the results are nearly perfect; when it does not, there will almost certainly be a detection error, which is at least predictable. The initialisation as it stands at the moment benefits from dense initial correspondence (based on the ordering of points) with ICP properly applied and its validity further verified. Unsurprisingly, all three datasets (GIP, NIST, Texas) are affected by this; without it, the PCA-GMDS approach would not work so well either (it just takes longer), as it heavily depends upon the finding of points between analogous points in almost every case. The built model would be without value unless there is consistency or contrariwise a reduction of irregular observations (robust PCA).

The observed behaviour is curious. While the process is somewhat stochastic and non-deterministic, upon different runtimes the results are usually more or less the same, maybe with a variation of a few percentile differences up or down, perhaps a fraction of 1%. But with particular pairs of surfaces there appears to be inconsistency as those same two surfaces are in a bit of a limbo. It is possible that GMDS will get the segmentation/correspondence wrong many times in a row (with widely varying stress values) and then ultimately get them right somehow. Figures $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , and $[*]$ give some examples of the process of debugging and the issues encountered along the way.

**Figure:** Initial experiments with the Texas3DFR Database excluded the cheeks, which were later added as various parameters were studied for their impact

**Figure:** Texas3DFR Database pairs with the correct correspondence

**Figure:** Model modes with more than 1% variation built from correct pairs

**Figure:** Model modes with more than 1% variation built from false pairs

**Figure:** With GMDS issues still in tact, the ROC curve for recognition suffers

**Figure:** A pair that GMDS usually fails on

**Figure:** Another pair that GMDS usually fails on

**Figure:** A closer, GMDS-style look on the very flawed correspondence-finding (example from Figure $[*]$ )

**Figure:** Another example of a GMDS-type comparison applied to a real pair and failing