This is an action detector for the Smart Classroom scenario. It is based on the RMNet backbone that includes depth-wise convolutions to reduce the amount of computations for the 3x3 convolution block. The first SSD head from 1/16 scale feature map has four clustered prior boxes and outputs detected persons (two class detector). The second SSD-based head predicts actions of the detected persons. Possible actions: raising hand and other.
|Detector AP (internal test set 2)||80.0%|
|Accuracy (internal test set 2)||90.5%|
|Pose coverage||Sitting, standing, raising hand|
|Support of occluded pedestrians||YES|
|Min pedestrian height||80 pixels (on 1080p)|
Average Precision (AP) is defined as an area under the precision/recall curve.
input, shape: [1x3x400x680] - An input image in the format [BxCxHxW], where:
Expected color order is BGR.
The net outputs four branches:
mbox_loc1/out/conv/flat, shape: [b, num_priors*4] - Box coordinates in SSD format
mbox_main_conf/out/conv/flat/softmax/flat, shape: [b, num_priors*2] - Detection confidences
mbox/priorbox, shape: [1, 2, num_priors*4] - Prior boxes in SSD format
out/anchor1, shape: [b, 2, h, w] - Action confidences
out/anchor2, shape: [b, 2, h, w] - Action confidences
out/anchor3, shape: [b, 2, h, w] - Action confidences
out/anchor4, shape: [b, 2, h, w] - Action confidences