Robust facial feature tracking using selected multi-resolution linear predictors

Sign Language Recognition

“…These approaches are more likely to be applicable to sign expressions, as they will have fewer constraints, having been trained on more natural data sets. An example of this is the work by Sheerman-Chase et al who combined static and dynamic features from tracked facial features (based on Ong's facial feature tracker [78]) to recognise more abstract facial expressions, such as 'Understanding' or 'Thinking' [89]. They note that their more complex dataset, while labelled, is still ambiguous in places due to the disagreement between human annotators.…”

4D Automatic Lip-Reading for Speaker's Face Identification

“…csamandhilda, that are consistent across a view, which indicates improved robustness to viewpoint updated for all the types of the primitive feature . The goal of this experiment is to obtain the best viewing angle for computing lip-reading and active appearance model (AAM) features that are extracted from each view, respectively by using their second derivatives .They use a linear predictor based on the tracking and it's has more robust lip-contour than the AAM that was introduced by [9].The audio-visual speech recognition system, visual features obtained from Discrete Cosine Transform( DCT) and active appearance model. ( AAM) were projected onto a 41 dimensional feature space using the LDA, proposed by [34] The systems reduce the dimensionality for Linear Discriminant Analysis (LDA) or Fisher's Linear Discriminant (FLD) as introduced by [57].…”

“…First, they used primitive way by combining the feature vectors (CFV), which called a cat and the second is concatenating the features and reduce the dimensionality using PCA, proposed by [14] which used (csam) feature. CFV improved by using an LDA over window for the set of frames .It is proposed by [15] which went to represented (hilda) features in the frontal lip-reading, It has applied for two features are the discriminating , it is introduced by [9] For all features, a z-score normalization is used, which has been visible to develop the separability between the features of the classes by [31] . The best viewing angle for the primitive features, i.e., those that aren't relative to a third PCA or an LDA i.e.…”

Improving recognition and identification of facial areas involved in Non-Verbal Communication by feature selection

“…Tracking was performed by linear predictor tracking [6]. Because the tracker requires multiple frames to be annotated for training, κ = 48 points {T i } i∈ [1..κ] that could be consistently located were selected for use and manually marked (see Figure 2).…”

“…The system uses linear predictor tracking [6] to track a selected set of facial locations, and makes use of geometric relations between points to encode facial shape information. Feature selection is then used to select the subset of feature components that are relevant to a specific NVC.…”