Evaluation code for COCO dataset

[lx709]

Dear Qi, your current code only provides the evaluation code for the FSOD dataset using 'Evaluation setting 1' as indicated in the supplementary of your paper. When will you release the code for getting the evaluation results on the COCO dataset? 'Evaluation setting 2' is more commonly used in current papers. Thanks for your attention.

[lx709]

and it would be nice if you can share your code for the few-shot fine-tuning process.

[1179021477]

me too

[fanq15]

I transferred the code on COCO dataset to detectron2 framework. I will release the detectron2 based code after the ECCV2020 camera-ready DDL.

[lx709]

Dear Qi, in my understanding, the main difference between evaluation setting 1 and setting 2 is that: in the former case, you randomly sample NK support instances to guide the detection of N10 query images for each episode, while in the latter case, we generally pre-select K support instances for each of the support classes (N_all=200), where N=5<N_all. If I understand correctly, evaluation setting 1 is more stable (the performance may vary a lot when the selected support varies in the evaluation setting 2) and relieves the number of support classes to a small number (N=5) rather than 200 in an all-way setting. At this point, evaluation setting 1 should be a better choice.

[fanq15]

Detectron2 based FSOD is released. It can reach 12AP on the COCO voc subset in the full-way 10-shot setting with ResNet50. The code is in https://github.com/fanq15/FewX.