1.3 KiB
Executable File
DISPENSION
INTOXIVISION PROJECT - YOLOV5-FACE
JANUARY 1, 2022
Lucas Wan (lucas.wan@dal.ca)
TO RUN
Ensure that all required packages are installed (see requirements.txt)
python3 detect_face.py --image "/image-location"
Can edit detect_face to update write location.
INFO
Uses pretrained model: yolov5m6_face. This model has the best recorded accuracy.
Landmarks output gives X Y coordinates of [Left Eye, Right Eye, Nose, Left Mouth, Right Mouth, Left Inner Eyebrow, Right Inner Eyebrow].
X = 0 is left of image (right = positive), Y = 0 is top of image (down = positive). X and Y range from [0 , 1].
Location of eyebrows are calculated from eye locations based on average distances between pupils (63mm) and between pupil to top of eyebrow (25mm).
Note that is folder only include files that are required for running the pretrained model (can not train a new model).
REFERENCES
https://github.com/ultralytics/yolov5
https://github.com/deepcam-cn/yolov5-face
https://www.techrxiv.org/articles/preprint/TFW_Annotated_Thermal_Faces_in_the_Wild_Dataset/17004538
TO DO
Combine landmark location information from multiple images (obtain average from burst of frames).
Identify central person (currently only outputting landmarks for 1 person - could be person off to the side).
Determine which packages in requirements.txt can be omitted.