YoungXIAO13 / FewShotDetection

(ECCV 2020) PyTorch implementation of paper "Few-Shot Object Detection and Viewpoint Estimation for Objects in the Wild"
http://imagine.enpc.fr/~xiaoy/FSDetView/
MIT License
212 stars 33 forks source link

3 times few-shot training instances? #23

Closed Lppy closed 3 years ago

Lppy commented 3 years ago

In meta dataset for VOC: https://github.com/YoungXIAO13/FewShotDetection/blob/master/lib/datasets/metadata.py#L39 and for COCO: https://github.com/YoungXIAO13/FewShotDetection/blob/master/lib/datasets/metadata_coco.py#L43

Variable shots are multiplied by 3:

if phase == 2:
    self.shots = shots * 3

where shots is the parameter passed by the command for training and is 10 or 30 for COCO: https://github.com/YoungXIAO13/FewShotDetection/blob/master/run/finetune_coco.sh#L6

and the variable self.shots are used as the number of instances for each class in function get_prndata: https://github.com/YoungXIAO13/FewShotDetection/blob/master/lib/datasets/metadata_coco.py#L234

Thus, in phase 2, you have 3 times instances compared with FSRW & TFA. Did I get it wrong?

YoungXIAO13 commented 3 years ago

Hi @Lppy

You can refer to issue#18 and issue#4, which pose a similar question.

Lppy commented 3 years ago

OK. Thanks for your reply.