EscVM / OIDv4_ToolKit

Download and visualize single or multiple classes from the huge Open Images v4 dataset
GNU General Public License v3.0
800 stars 633 forks source link

Annotation bounding box coordinates with invalid values? #33

Closed monocongo closed 5 years ago

monocongo commented 5 years ago

I have downloaded images and label files for an image class, and I'm wondering if maybe I've done something to monkey with the bounding box coordinates in the label *.txt files.

For example, I have a label file (0a7df07bbac03159.txt) with the following contents:

Sword 14.72 7.68 669.44 767.360256

My understanding from the README is that the annotation bounding box coordinates should be within the normalized range [0, 1], but that's obviously not what I'm getting in my label files, as seen above.

Can anyone comment as to how I should interpret the bounding box values in the *.tx files, and/or what I've done wrong to get values that appear to be outside the expected range? Are the float values present in these files computed from the original/normalized bounding box values against the height/width of the corresponding image, and if I want to have the actual integer pixel numbers for the bounding box then I can just round these to the nearest integer? For example, the bounding box for the above would be (left_x=15, top_y=8, right_x=669, bottom_y=767)?

Thanks in advance for any comments or suggestions.

EscVM commented 5 years ago

Hi @monocongo!

You are right, the description is not anymore correct. Those are the values computed from the normalized images and so, as you wrote, the corresponding rounded values are (left_x=15, top_y=8, right_x=669, bottom_y=767).

monocongo commented 5 years ago

Thanks @EscVM !