Ongoing Progress and Results

T he resultant models and their performance when used in generative mode were the initial results we got for the aforementioned framework. But over time we moved on to exploring other areas, which also fall under this massive section, for reasons of convenient.

Having overcome most of the barriers associated with pre-processing of GIP data, models were built using a couple of separate sets, each containing a sequence of expressions of a certain type. These two were compared in the discrepancy sense, which in turn was modeled by PCA and yielded Figure $[*]$ (experiment from 9/4/2011 with nose tip search near the centre, with smoothing, expression of surprise compared to sadness).

**Figure:** Decomposition based on sample GIP data (Pareto)

**Figure:** Decomposition based on just 4 registered faces with different expressions (Pareto)

Values of 0.0018746, 0.0016749 and 0.002639, accounting for 38.9452, 34.7974, and 26.2574 percent of the observed variation.

We have run experiments on the largest expression datasets we have, weighing 440MB and 880MB (see Figure $[*]$ ). This algorithm works a lot more reliably now, however matching/scoring remains to be done. We expressions-free datasets to register to?

**Figure:** Same as prior figures, but with 90 images in the set

Next, matching criteria based on the model will be used to score for recognition rates; larger experiments can then be engineered to produce ROC curves. In order for such large experiments not to require reruns, however, it will be desirable to further refine the existing implementation while improving the GUI, the documentation, and also ensuring the route taken is widely accepted and not an enormous effort embarked on in vain.

Figure $[*]$ shows the percentage of the explained variation for sets of 9 images; ``fear'' to ``surprise'' yields 24.0210, 18.9342, 13.2081, 12.6385, 10.1781, 9.1887, 7.0255, and 4.8060 percent, corresponding to the magnitude of the 8 modes of variation.

**Figure:** Overview of an experiment dealing with expression-to-expression variation model-building

At the point where we are prepared to run larger, systematic experiments there might be use for GIP's implementation of ICP. The detection has come to the point where it can quite reliably capture the right parts, at least based on some early observations.

All control of probe-to-gallery experiments was moved to the GUI side such that iteration will be simpler and require less manual work or future coding. The implementation was made somewhat more elegant by merging similar bits of code and ensuring deprecated parts are removed or hidden away in secondary options. Assuming that the data is reasonably clean by now¹⁶ (for GIP data it is a lot more manageable now), we are ready to run experiments and then vary parameters to extend these to ROC curves, as planned. EDMs can be constructed for pairs of expressions or neutral/non-neutral. The 3-D portion of GC data is said to contain just smile shots and neutral shots for each subject, so perhaps for these experimental results will be dependent upon whether the model is trained with just smiles or all sorts of different expressions. Perhaps the GC models should be treated in total isolation from GIP ones, as the curves adhere to a certain standard which is independent of locally-acquired data.

GIP data contains data from one subject (at least for expressions), so coming up with a way to test recognition of a one subject among many (inter- or cross-personal) is a non-starter. Preliminary results ought to show only feasibility, so these will be run as a side task while we return to GC data for experimental validation.