Dr. Philip A. Tresadern

Face Alignment Resource

Face recognition has long benefitted from established evaluation protocols: all methods assess their performance using the same data as each other, following an agreed procedure. Face alignment (i.e. localizing individual features on the face, such as the eyes, nose and mouth) is missing such a protocol, such that it is difficult to compare methods.

As a first step to addressing this problem, we propose a simple web resource that contains, for several standard datasets, true feature locations and several sets of perturbed locations. These perturbed locations can be used for training (e.g. an Active Appearance Model) and, more importantly, for testing. If you have developed and tested a new algorithm using these precomputed initial points, we encourage you to send us your outputs (i.e. the estimated feature locations) so that they can be shared with the community.

This resource will have several benefits:

• Improved productivity: Implementing someone else's method is hard, especially if the method were published in a short conference paper. Having easy access to the raw outputs (rather than derived statistics) makes it much easier for you to compare your method with the current state of the art.
• Fair comparison: By ensuring that all methods start from the same sets of initial points, we can be more confident that differences in performance between two methods are not an artefact of randomness in initialization.
• Avoiding poor implementations: Because implementing someone else's method is hard, it is unlikely that someone else's implementation of your method will be as effective as your own implementation. It is therefore likely that your method will compare unfavourably in their study, and there is a risk that an inferior method is published due to a poor comparison. Providing access to your results will ensure that your method receives fair treatment in other studies.

Supporting Documents

For transparency, we provide some example matlab scripts that we used to generate the perturbed points from their hand-labelled locations. These scripts also include code for computing statistics from the estimated point locations and generating cumulative frequency curves for use in publications. Example statistics and figures are included in the results for Tresadern et al., IJCV 2011.

We also provide images of the markup scheme for a 68 point model [png] [pdf], 22 point model [png] [pdf] and 20 point model [png] [pdf], and the average shape model in the normalized model frame [68 points] [22 points] [20 points].

The 22 and 20 point models are derived from subsets of the 68 point model [68->22] [68->20].

Results: Medium Perturbation