Bounding box about space HOT 8 CLOSED

Hi @bentennyson9x, z_where here refers to the bounding box. Its shape is (B, G*G, 4), where G*G is the number of cells (e.g., 4*4). z_where[:, :, :2] gives the size of the object (normalized to (0, 1)) and z_where[:, :, 2:] gives location (normalized to (-1, 1)). Since many of the cells won't contain valid objects, you also need z_pres here to determine which cells contain objects.

To convert z_where and z_pres (you should binarize z_pres before sending it in) into list of bounding boxes like [[(y_min, y_max, x_min, x_max)]] in pixel values, you can use the convert_to_bbox function here.

from space.

Thanks for your help! I got it

from space.

I got an error "IndexError: The shape of the mask [256, 1] at index 1 does not match the shape of the indexed tensor [256, 4] at index 1" when I passed convert_to_boxes(log.z_where,log.z_pres, log.z_pres_prob) to the "convert_to_boxes" function. I checked again and realized that log.z_pres always has a shape (256,1) (because sigmoid) while log.z_where shape is ([256, 4] with [sx, sy, tx, ty]. Am I missing something?

from space.

Ah sorry I forgot to mention this thing: you should squeeze z_pres and z_pres_prob from (B, N, 1) to (B, N) (i.e. z_pres.squeeze(2)) before sending them to convert_to_boxes as shown here

from space.

I got bad values of the bounding box. I think it is due to the binarizing z_pres in 0 and 1 value. I don't know the threshold to do it. Can you help me. z_pres values are in this picture when I print it out

from space.

A simple z_pres = z_pres > 0.5 should do. How did you binarize it? The result is supposed to be ByteTensor but what you show me seems to be FloatTensor.

from space.

Oh, I did it! Thank you very much. You save my day!

from space.

You're welcome :)

from space.