TH
Hall of Faces: a face detection model zoo
A collection of face detection models pre-trained on the Widerface
dataset.
In the table below you can see each model detailed information including:
- meta architecture name
- model speed
- detector performance measured on the FDDB benchmark
- a download link to a
tar.gzfile containing the model and configuration files - a link for a live demo running on a Google Colaboratory notebook
| Architecture | Speed (ms) | mAP@0.5 | Cfg/Weights | Demo |
|---|---|---|---|---|
| R-FCN resnet101 | 92 | 94.73 | link | colab |
| Faster R-CNN inception resnet v2 atrous | 620 | 94.39 | link | colab |
| SSD mobilenet v1 | 30 | 91.20 | link | colab |
| YOLOv2 | 15 | 89.59 | link | colab |
| TinyYolo | 5 | 85.5 | link | colab |
Face detectors performance evaluation on the FDDB dataset
Discrete ROC
Continuous ROC
Training details
Morghulis was used to
download and convert it to either Darknet or Tensorflow Object Detection API format.
Tensorflow Object Detection API
The remaining models were trained with Tensorflow Object Detection API
on Google Cloud ML Engine.
Darknet
There are 2 models trained with Darknet: one based on YOLOv2 and other
on Tiny YOLO. Both used convolutional weights that are pre-trained on Imagenet:
darknet19_448.conv.23.