Face-recognition systems usually work under the assumption that first of all landmarks, such as eyes, nose, and mouth, need to be found in an image accurately. These are then used to map model faces to the image in a topographically constrained way, i.e. the geometry of the model faces has to be preserved, usually with some dostortion tolerance. Only then can recognition be achieved.
This project investigates how well a system can perform, if no landmarks in this sense are used and if the matching of local features on the image is not topographically constrained, or at least only in a very crude way. Models and images are represented by regular and sparse grids of jets, 10×10 for models and 16×17 for images, which are larger. The similarities between model jets and image jets can be written in a large similarity matrix of 100×272. For a given image there is one similarity matrix for each model.
The sum over the similarity matrix performed very poorly, but taking the maximum similarity for each model node, which is nearly as simple as taking the sum, performed as well as our system for Face Recognition by Dynamic Link Matching. Moreover, if some topographical constraints were taken into account in a very simple way, the performance was comparable to that of Face Recognition by Elastic Bunch Graph Matching, at least for frontal views. If rotation in depth was 22°, elastic bunch graph matching performed significantly better than the simple method. Some other aspects of the systems were considered as well and it was confirmed, for example, that recognition performance is best if matching is done with Gabor-phase (and recognition is done without phase).
|model pose||image pose||gallery size||elastic bunch graph matching||simple system with some topographical constraints||simple system with no topographical constraints||gallery size||dynamic link matching|
|frontal||frontal||108||91 %||92 %||88 %||111||85 %|
|frontal||11°||108||94 %||94 %||91 %||111||92 %|
|frontal||22°||108||88 %||81 %||66 %||111||66 %|
Table: First rank recognition rates on the Bochum database for different systems.