Translation Explored

Encouraging results come about with the changes to the ICP routines, adding to what previously was tested at another level, namely model and residual. With improved smoothing and with GIP's v1 of ICP (older) the results are improved significantly. Visual examples with and without translation (but with GIP's v2 of ICP) are shown...

Improving this further with rotation and a model that annuls translation and rotation (currently it does not) should be trivial, with additional refinements incurred by use of larger models and simpler sets (the ones currently used are full of different expressions).

**Figure:** Example differences between an image before and after translation (in all three dimensions)

**Figure:** On the left: the results from a test run (first 20 images) using the determinant-based objective function. The model was not constructed with translation, whereas matching did. On the right: the same but with a median-based similarity measure.

**Figure:** A top-to-bottom view of one's face (the rigid part) with corresponding translation and rotation

**Figure:** A look at some of the tweaking and debugging process of ICP, where the angle shown is pointing from underneath the nose, going towards the top

**Figure:** An example of the first image pair, visualised separately as stripes as coarse as the image sampling rate (for the model)

ICP experiments were run until papers were explored again. One experiment in the pipeline strives to emulate ICP as described in Mian's earlier papers, on which he based his Ph.D. (he did it at the same time as myself).

Upon completion of a larger experiment it turned out that it was designed incorrectly because ICP - with Mian-style translation¹⁸ - did not work correctly and therefore the model built was improper. It did occur, however, that despite this fluke there is decent capability within the new model to detect pairs (maybe an accidental discovery worth exploring in the future). The results are shown in Figure $[*]$ .

Our newly-combined set of Spring and Fall Semester builds a model with translation, where the ICP algorithm is as the same as UWA's. The test sets are also of Spring and Fall Semester, but they have no overlap with respect to the training set and they still have many expressions in them. To distinguish between expression differences (intra-personal) and inter-personal differences we use a coarse model which is fast to build and match to.

Unintended arrival at experiments where a model-based approach by far outperforms the median-based are probably not of high priority at this stage.

**Figure:** The results of a mis-constructed experiment where ICP did not work correctly and nonetheless, the model-based approach did not fail so miserably

To begin the exploration of robust PCA as proposed by Yi Ma (of MSR- China, a copy of his book (350+ pages) was obtained, hopefully with concise summaries too.