Hi all, I have an idea I want to try on the OAK-D but not sure how to approach it. I want to feed a cropped and resized frame into another neural net from a face detect net but avoid the host. I assume this'll be much faster. From [here](https://rahulrav.com/blog/depthai_camera.html) it seems like a custom PyTorch net in between could do it. The problem is I would have to link the camera preview and the face detect bounding box output both to this new net. Is that possible? What would the data format be? I checked the docs but couldn't find any leads.

Linking two nodes as input to a neural net

naldic

Hi all,
I have an idea I want to try on the OAK-D but not sure how to approach it. I want to feed a cropped and resized frame into another neural net from a face detect net but avoid the host. I assume this'll be much faster.
From here it seems like a custom PyTorch net in between could do it.

The problem is I would have to link the camera preview and the face detect bounding box output both to this new net. Is that possible? What would the data format be? I checked the docs but couldn't find any leads.

erik

Hello naldic,
from my understanding you would like to first detect the face, crop the original image and feed that (just the face) to the second-stage NN (for example face landmark/emotion/age-gender detection NN)? I am just in the process of refactoring gen2-age-gender experiment to do just that - cropping on the device based on the face detecton of the first NN using script node. It should be refactored today and hopefuly I will remember to circle back with the PR (of the updated experiment).
Thanks, Erik

naldic

Oh great that looks like what I need. I'll try that approach thanks.

Brandon

https://github.com/luxonis/depthai-experiments/pull/169