I run the face recognition demo according to README.md in the directory /home/<user>/intel/openvino/deployment_tools/open_model_zoo/demos/python_demos/face_recognition_demo, but can't recognize the face. The terminal commands executed are as follows:
python3 face_recognition_demo.py -m_fd /home/<modelPath>/face-detection-retail-0004.xml -m_lm /home/<modelPath>/landmarks-regression-retail-0009.xml -m_reid /home/<modelPath>/face-reidentification-retail-0095.xml -l /home/<user>/intel/openvino/deployment_tools/inference_engine/lib/intel64/libcpu_extension_sse4.so --verbose -fg "/home/<user>/data"
Part of the log printed in the terminal is as follows:
[ DEBUG ] 2019-12-11 10:33:07,917 Adding label liu-3 to the gallery. [ DEBUG ] 2019-12-11 10:33:07,917 Adding label liu to the database [ DEBUG ] 2019-12-11 10:33:07,917 The database length is 1. [ DEBUG ] 2019-12-11 10:33:07,942 Adding label liu-2 to the gallery. [ DEBUG ] 2019-12-11 10:33:07,942 Appending new descriptor for label liu. [ DEBUG ] 2019-12-11 10:33:07,943 The database length is 1. [ DEBUG ] 2019-12-11 10:33:07,972 Adding label fei to the gallery. [ DEBUG ] 2019-12-11 10:33:07,973 Adding label fei to the database [ DEBUG ] 2019-12-11 10:33:07,973 The database length is 2. [ DEBUG ] 2019-12-11 10:33:08,004 Adding label liu-1 to the gallery. [ DEBUG ] 2019-12-11 10:33:08,004 Appending new descriptor for label liu. [ DEBUG ] 2019-12-11 10:33:08,004 The database length is 2. [ INFO ] 2019-12-11 10:33:08,004 Database is built, registered 2 identities
Database creation and model input are correct, why the text above the face bounding box is always unknown?
Can anyone help?
I'm not sure how one can use this face recognition model. I had to bump the t_id to .5 before the unknowns showed some labels, but half the time they were wrong. I'm not sure how you can get a good recognition with 5 landmarks. Anyone able to use a 35 or 68 landmark model on this with better success?
I finally started seeing name labels once I made t_id=.5
However, the accuracy is very poor. I'm not sure how you can get reasonable good accuracy with 5 landmark detections.
Does anyone have any success with using 35 or 68 point landmark models? If so, could you share?