Annotations

Many databases come with additional information about their data. For image databases, e.g., the locations of hand-labeled facial landmarks are provided. Usually, these data is stored in additional text files. For most of the available bob.db databases, there is exactly one text file for each data file.

The function bob.db.base.read_annotation_file() can be used to read annotation files of different types. It will output the data as a dictionary, containing a key and the interpreted read data. For landmark locations, the data is returned in the common way for bob, which is (y, x)! The following formats are currently accepted:

• 'eyecenter' (for face images): Each file contains only the locations of the two eyes, in one row, as follows: re_x re_y le_x le_y. The keys will be 'reye' and 'leye'.

• 'named' (for face images): Each file contains lines with the landmark name and the two landmark locations, e.g. reye re_x re_y.

• 'idiap' (for face images): The file format to read Idiap specific annotation files. It will return up to 24 key points. 22 of these are read from the file, and the 'reye' and 'leye' are estimated from the inner and outer corners of the eyes (if available).

Note

Left and Right positions are always expected to be from the subject perspective. This means that, e.g., the 'leye' landmark usually has a higher x-coordinate than the 'reye'.