SRILF Examples

Back to All Examples

 

SRILF Examples with occlusions and rotations

(Bosphorus database)

In this experiment we target 14 landmarks on 4339 facial scans from the Bosphorus database, training the model using 2-fold cross validation. The examples below focus on scans with occlusions and out-of-plane rotations. However, training was done exclusively with frontal scans without occlusions, so that the model could not learn from the artifacts present in occluded or rotated scans. The examples below are divided in 4 categories:

·        Scans with occlusions due to hair, eye glasses or hands.

·        Scans with moderate rotations.

·        Scans with large rotations: in these cases, large parts of the surface become self-occluded. To make the results clearer, we show two views for each of these scans: the original one and a rotated one so that the scan is viewed from an approximately frontal viewpoint.

·        Scans that are especially challenging due to poor quality and/or extreme occlusions.

Automatically located landmarks are indicated by small blue or red spheres, to distinguish whether their position was determine based on candidate points provided by the feature detectors (blue) or by inference from the statistical shape model (red).

Scans with occlusions

166

299

456

977

1332

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Scans with moderate rotations

166

299

456

977

1332

 

 

 

 

 

 

 

 

 

 

Scans with large rotations (2 views for each scan)

 

Imagen1.png

Imagen2.png

Imagen3.png

166

299

456

977

1332

Imagen4.png

Imagen5.png

Imagen6.png

 

 

 

 

 

Imagen7.png

Imagen8.png

Imagen9.png

 

 

 

 

 

 

Especially challenging scans

166

299

456

977

1332

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

All Examples