I am not able to understand the output format of the person-detection-action-recognition-0006 model. It says there is 8600 number of priors but actually, when I tried to run on the local machine I only got 8550, 50 are missing. Along with that, I am not able to detect the actions/bounding box as mentioned in the documentation.
The documentation says -
num_priors - number of priors in SSD format (equal to 50x86x1+25x43x4=8600)
How did you compute num_priors as 8550? Can you elaborate?
A recommendation is to check the respective values during runtime by adding the watch on the variables.
Priors are pre-computed, fixed-size bounding boxes that match the ground truth.
I used the following command to verify the sample:-
./smart_classroom_demo -i classroom.mp4 -m_act intel/person-detection-action-recognition-0006/INT8/person-detection-action-recognition-0006.xml -m_fd intel/face-detection-adas-0001/INT8/face-detection-adas-0001.xml -m_lm intel/landmarks-regression-retail-0009/FP32/landmarks-regression-retail-0009.xml -m_reid intel/face-reidentification-retail-0095/FP32/face-reidentification-retail-0095.xml -l lib/libcpu_extension.so -al out.txt
Sample Output in out.txt file:-
rect: [ 446, 488, 178, 395 ]
According the reported network, it has two heads and 5 predefined “prior boxes”:
There is a difference between the terms “prior box” and “anchor box”.