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 1/8 and 1/16 scale feature maps 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: sitting, writing, raising hand, standing, turned around, lie on the desk.
|Detector AP (internal test set 2)||90.70%|
|Accuracy (internal test set 2)||80.74%|
|Pose coverage||sitting, writing, raising_hand, standing,|
|turned around, lie on the desk|
|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: [1x400x680x3] - An input image in the format [BxHxWxC], where:
Expected color order is BGR.
The net outputs four branches:
ActionNet/out_detection_loc, shape: [b, num_priors*4] - Box coordinates in SSD format
ActionNet/out_detection_conf, shape: [b, num_priors*2] - Detection confidences
ActionNet/action_heads/out_head_1_anchor_1, shape: [b, 6, 50, 86] - Action confidences
ActionNet/action_heads/out_head_2_anchor_1, shape: [b, 6, 25, 43] - Action confidences
ActionNet/action_heads/out_head_2_anchor_2, shape: [b, 6, 25, 43] - Action confidences
ActionNet/action_heads/out_head_2_anchor_3, shape: [b, 6, 25, 43] - Action confidences
ActionNet/action_heads/out_head_2_anchor_4, shape: [b, 6, 25, 43] - Action confidences
[*] Other names and brands may be claimed as the property of others.