Detection ground truths don't match the coco17 validation ground truths

[Britefury]

Hi,

I have been taking a look at the ground truths that you've provided for the detection competition. I have noticed that the GT boxes in the val_ground_truth.pkl do not match those in coco17-val.txt; it seems that they have been scaled. A consistent x,y scale factor is used for each image, but I am unable to determine the pattern/algorithm used to compute the scale factors. Would it be possible to replace val_ground_truth.pkl with a file generated by direct conversion?

I have created a pull request (#11) that adds a conversion script (in case you're interested) and also changes the detection README slight in order to clarify the first problem that I had concerning format :)

[Britefury]

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.

[MInner]

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.

[Britefury]

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.

[Britefury]

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.