openpose/doc/standalone_face_or_hand_keypoint_detector.md

OpenPose Library - Standalone Face Or Hand Keypoint Detector

In case of hand camera views at which the hands are visible but not the rest of the body, or if you do not need the body keypoint detector and want to speed up the process, you can use the OpenPose face or hand keypoint detectors with your own face or hand detectors, rather than using the body keypoint detector as initial detector for those.

OpenCV-based Face Keypoint Detector

Note that this method will be faster than the current system if there is few people in the image, but it is also much less accurate (OpenCV face detector only works with big and frontal faces, while OpenPose works with more scales and face rotations).

./build/examples/openpose/openpose.bin --body 0 --face --face_detector 1

Custom Standalone Face or Hand Keypoint Detector

Check the examples in examples/tutorial_api_cpp/, in particular examples/tutorial_api_cpp/06_face_from_image.cpp and examples/tutorial_api_cpp/07_hand_from_image.cpp. The provide examples of face and/or hand keypoint detection given a known bounding box or rectangle for the face and/or hand locations. These examples are equivalent to use the following flags:

# Face
examples/tutorial_api_cpp/06_face_from_image.cpp --body 0 --face --face_detector 2
# Hands
examples/tutorial_api_cpp/07_hand_from_image.cpp --body 0 --hand --hand_detector 2

Note: both FaceExtractor and HandExtractor classes requires as input squared rectangles.

Advance solution: If you wanna use the whole OpenPose framework, you can use the synchronous examples of the tutorial_api_cpp folder with the configuration used for examples/tutorial_api_cpp/06_face_from_image.cpp and examples/tutorial_api_cpp/07_hand_from_image.cpp.

Cropping the Image for Hand/Face Keypoint Detection

If you are using your own hand or face images, you should leave about 10-20% margin between the end of the hand/face and the sides (left, top, right, bottom) of the image. We trained with that configuration, so it should be the ideal one for maximizing detection.

We did not use any solid-color-based padding, we simply cropped from the whole image. Thus, if you can, use the image rather than adding a color-based padding. Otherwise black padding should work good.