Test result of COCO dataset

[dhgras]

I tried to reproduce according to the steps in the README, and finally got the test results as shown in the figure.
I don't know whether this result is correct. If not, where did I go wrong? And what else do I need to do?

[Jaraxxus-Me]

1. If you're testing with the ResNet101 backbone model, 3-shot setting, I think the result may be wrong. I remember the result on the first image is not like that.
2. You will need to set a proper threshold for better visualization, I usually use 0.8~0.9 for the images displayed in the paper. For the calculation of AP scores, we follow a common setting as MaxDet=100.
3. If you notice the scores of the models on COCO, the mAP is still less then 10, which means few-shot detection w/o fine-tuning is still super challenging in COCO (there are still many false postives). So it is common to find failure cases and we're looking forward to its future development.