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.
name: "input" , shape: [1x3x384x672] - An input image in the format [BxCxHxW], where:
Expected color order is BGR.
image_id
, label
, conf
, x_min
, y_min
, x_max
, y_max
]image_id
- ID of the image in the batchlabel
- predicted class IDconf
- confidence for the predicted classx_min
, y_min
) - coordinates of the top left bounding box cornerx_max
, y_max
) - coordinates of the bottom right bounding box corner.[*] Other names and brands may be claimed as the property of others.
The NET was tuned from face-detection-adas-0001 weights