Detection ground truths don't match the coco17 validation ground truths about visda-2018-public HOT 4 OPEN

Further investigation leads me to conclude that most of the Coco image ground truths were scaled to a resolution of 300x300, with some being scaled to either 300x150 or 150x300.
I can get around this by computing the per-image scale factor and applying it my predictions, but this won't work at the test stage since I wont have the ground truths from which to compute the scale factors.

from visda-2018-public.

Thanks a lot for this finding and the pull request! Let me take a closer look. That is weird because I remember specifically addressing this issue at some point.

from visda-2018-public.

Okay, I've managed to make a submission to CodaLab and achieve a score that beats simple supervised training.

In order to do this though, I had to compare val_ground_truth.pkl in the repo and coco17-val.txt, compute the per-image bounding box scale factor and apply them to the predictions from my algorithm before submitting them, otherwise I get a mAP score of around 0.1% / 0.001.

from visda-2018-public.

I've added a script to the PR that compares the GT boxes in coco17-val.txt with those in val_ground_truth.pkl and prints the per-image scale factor. If you run it, you will notice that a scale factor != 1 is applied to the boxes for every image.

from visda-2018-public.