Table 4 The normalised mean error of landmark detection for the ELD model, trained on different subsets from DogFLW, is evaluated on the test set. In the case of “included,” all breeds were present in the training set. Conversely, in the “excluded” case, the selected breed was absent from the training set. The error is reported for both the full test set and for the selected breed within the test set.

From: Dog facial landmarks detection and its applications for facial analysis