Problems with detect large objects? about detr HOT 5 CLOSED

Oh, yeah, I will retest to check this bug, thanks for your guys quick response.

from detr.

Hi,

What is the input image resolution that you are feeding to your model? Note that DETR has been trained on images with min size of 800 and max size of 1333.

Also, which model are you using, detr_resnet50 from the torchhub, or the simplified demo model from the colab?

from detr.

Hi, I just ran your image (well, together with bounding boxes since I don't have the original) with the demo colab code and got this, there are predictions over what's already been on the image but large predictions are there

from detr.

So, thanks a lot @jinfagang for opening the issue.

It turns out that there is a bug in the visualization -- we are using itertools.chain to generate the infinite colors for the boxes, but it actually stops once the initial set of colors is used.

We are fixing this in the colab, but by replacing

    colors = itertools.chain(COLORS)

with

    colors = COLORS * 100

we get (after bumping the threshold a bit as well)

from detr.

Given that I believe we addressed the original issue, I'm closing this, but let us know if you have follow-up questions

from detr.