1
$\begingroup$

I have rectangular bounding boxes for an object in an image, and 3-point vector of relative distances (x,y,z) of the object w.r.t camera, and 3-point vector indicating the pose of the object. I want to take the cropped patch (from bounding boxes) and resize them to 224 X 224 and pass it through to some standard networks like resnet (training from scratch) to estimate pose. Now, I think results can be improved if I use the distances and the aspect ratio (from bounding box coordinates) as input too. How do I go about doing that?

$\endgroup$

1 Answer 1

1
$\begingroup$

You could create a network which merges the output layer of a convolution network like ResNet (minus the top layer, if you are using pre-trained weights), and an input layer that feeds your additional inputs (the distances and aspect ratios).

You can experiment with the type of merge operation, concatenation should be a good starting point. There is a basic example of the operation, using Keras, here.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.