DogFLW: Dog Facial Landmarks in the Wild Dataset
The Dog Facial Landmarks in the Wild (DogFLW) dataset offers 3,274 images of dogs' faces in various environments and conditions, each annotated with 46 facial landmarks and a bounding box on the dogs' face. This dataset is inspired by the CatFLW (Cat Facial Landmarks in the Wild set) and designed to aid researchers in analyzing dog emotions and behaviors.
The development of the landmark scheme was carried out by experts certified in DogFACS coding. It enables detailed facial expressions, guided by the anatomy of dogs' face. The result is a detailed set of 46 points capturing essential facial features, including the nose, eyes, mouth, and ears.
THE LANDMARKS
brow left
ear left upper intersection with head
ear left upper third
lip l corner
lip r upper mid
ear right upper intersection with head
lip r corner
ear r lower intersection with head
eye r outer
lip lower middle
chin
ear left lower intersection with head
tongue tip
brow right
eye r upper
ear left lower third
eye r lower
eye r inner
zygomus right
zygomus left
ear left tip
ear right tip
nose bottom
nostrils r outer
ear r upper 2thirds
nose upper
ear r upper third
ear r lower third
ear r lower 2thirds
whiskerspad r mid
snout right
eye l upper
eye l outer
eye l lower
eye l inner
snout left
nosepad mid
nose r upperedge
nostrils middle
nose l upperedge
nostrils r outer
lip l upper mid
lip upper middle
whiskerspad l mid
ear r lower 2thirds
ear r upper 2thirds
The dataset draws from the Stanford Dog dataset and includes individual dogs from 120 different breeds. The images are varied in resolution, and are set in non-laboratory conditions, capturing real-world environments.
To evaluate the dataset, we used the Ensemble Landmark Detector (ELD) model, which provided benchmark results using normalized mean error (NME) as a performance metric. The results show high accuracy in detecting landmarks around the eyes and nose but present challenges with ears due to their variability in shape and position across breeds. Despite these challenges, the DogFLW dataset serves as a robust resource for understanding dog facial expressions. For detailed information, please see DogFLW: Dog Facial Landmarks in the Wild Dataset paper.
We believe that animal affective computing crucially depends on the availability of openly available annotated datasets. We are proud to lay the foundations for automated analysis of animal facial expressions, and we invite the scientific community to help us develop this direction.
​
The dataset creation was generously supported by the Data Science Research Center (DSRC) of the University of Haifa.
Project Team
COLLABORATORS