Mut1ny’s fast head detector is a deep neural network human head detector that has remarkable fast detection speed. The Head detector has a wider range of application than your typical face detector. Face detectors as the name already implies can only detects faces of people. This means certain viewing angles towards the camera they can not cover. In particular back and extreme side facing heads or bird eye views. Our head detector can cover the full hemisphere of people’s heads towards the camera.
Additionally it covers a bigger spatial area. Often face detector really only cover the face and therefore tend to crop interesting areas like ears and parts of the hair. Mut1ny’s fast head detector is able to cover both and sometimes even the full area of hair.
The detectors network architecture is based on Version five in the popular YOLO network architecture family. The network has been trained on our unique head dataset that covers a wide range of scenarios, different ethnicity, gender and age .
This product is the ideal companion in combination with our other facial processing networks, especially if the head only takes up a smaller percentage of the image/frame. Or if there are multiple faces in the footage.
The model is deliverable in these supported neural network formats
But if you need any other neural network format just get in contact with us.
What is included in the fast head detector deep neural network?
- The neural network model in the formats above
- For PyTorch format only:
- Python inference code using our model
- C++ inference code using our model
- For ONNX format only:
Why buy a deep neural network model? When you can build and train one yourself? Or get one free from the internet?
- Training a segmentation model from scratch using a large enough dataset takes between 2-3 days even with a high-end equipped GPU. This translates to € 80-150,- public cloud instance spending costs alone.
- There is no need to spend any development time, but instead you can fully concentrate on your product feature.
- Acquiring a large enough annotated dataset can take man-months.
- To our knowledge there are no public head detection models available that cover a similar range in gender, ethnicity, age & pose. Those that do exist most likely have a very restrictive licensing scheme.