Use Case and High-Level Description

Face detector for driver monitoring and similar scenarios. The network features a pruned MobileNet backbone that includes depth-wise convolutions to reduce the amount of computation for the 3x3 convolution block. Also some 1x1 convolutions are binary that can be implemented using effective binary XNOR+POPCOUNT approach




Metric Value
AP (head height >10px) 31.2%
AP (head height >32px) 76.2%
AP (head height >64px) 90.3%
AP (head height >100px) 91.9%
Min head size 90x90 pixels on 1080p
GFlops 0.611
GI1ops 2.224
MParams 1.053
Source framework PyTorch*

Average Precision (AP) is defined as an area under the precision/recall curve. Numbers are on Wider Face validation subset.



  1. name: "input" , shape: [1x3x384x672] - An input image in the format [BxCxHxW], where:

    • B - batch size
    • C - number of channels
    • H - image height
    • W - image width

    Expected color order is BGR.


  1. The net outputs blob with shape: [1, 1, N, 7], where N is the number of detected bounding boxes. For each detection, the description has the format: [image_id, label, conf, x_min, y_min, x_max, y_max]
    • image_id - ID of the image in the batch
    • label - predicted class ID
    • conf - confidence for the predicted class
    • (x_min, y_min) - coordinates of the top left bounding box corner
    • (x_max, y_max) - coordinates of the bottom right bounding box corner.

Legal Information

[*] Other names and brands may be claimed as the property of others.

The NET was tuned from face-detection-adas-0001 weights