A question about the input and target about dsntnn HOT 4 CLOSED

raccoon_face_tensor = torch.from_numpy(raccoon_face).permute(2, 0, 1).float()

input_tensor = raccoon_face_tensor.div(255).unsqueeze(0)

input_var = input_tensor.cuda()

the input_var is normalized (0~1).

But I don't know the target_var.

eye_coords_tensor = torch.Tensor([[[eye_x, eye_y]]])

target_tensor = (eye_coords_tensor * 2 + 1) / torch.Tensor(image_size) - 1

target_var = target_tensor.cuda()

If I want to use DSTNN, what do I need to do with the input and output?

from dsntnn.

This library offers the ability to choose whether you would like to calculate normalised coordinates or pixel coordinates via the normalized_coordinates parameter of the dsnt function. If you leave this as its default (False), then you will need to normalise the target. You can convert between normalised and pixel coordinates using the normalized_to_pixel_coordinates and pixel_to_normalized_coordinates functions.

from dsntnn.

I would like to ask, if I use DSTNN to calculate coordinates, do I need to do any standardization on my input and the heat map label?

At present, my image is standardized with [128,128, 128] as the mean value, [256,256,256] as the variance, and the label is the Gaussian distribution of the key points in the image position.

from dsntnn.

The heatmap which goes into dsnt is expected to be a valid probability distribution over locations, so all pixels must be be non-negative and sum to 1. You can achieve this using flat_softmax on the network's output. Please refer to the basic usage guide for an example. You do not need to create the Gaussian distribution for your targets, the functions in this library take 2D keypoint locations.

from dsntnn.