search

hzhupku / DCNet

Dense Relation Distillation with Context-aware Aggregation for Few-Shot Object Detection, CVPR 2021
MIT License
142 stars 32 forks source link

plz check your code #7

Closed Wei-i closed 2 years ago

Wei-i commented 2 years ago

https://github.com/hzhupku/DCNet/blob/a73f448fc664d8c5aabf8b7b459942e8c576e6a6/maskrcnn_benchmark/modeling/detector/generalized_rcnn.py#L47

Wei-i commented 2 years ago

https://github.com/hzhupku/DCNet/blob/a73f448fc664d8c5aabf8b7b459942e8c576e6a6/maskrcnn_benchmark/data/datasets/__init__.py#L9 comment out this line

Wei-i commented 2 years ago

Another issue made me confused is that how I can new a DCNet/fs_list/voc_traindict_bbox_Xshot.txt and DCNet/fs_list/voc_traindict_full.txt instead of using the files you provide in which each line contains the absolute path. When preparing voc_dataset, the dir is 'datasets/VOCdevkit/..', however when training, the root_dir is datasets/voc/VOC{year} and so on . BTW, these made me feel sad to reproduce by my poor ability.

Wei-i commented 2 years ago

voc_split_3_base_training:

421ef91b4147013b422cdbfcb1077fd
Wei-i commented 2 years ago

reproduce reuslt: voc_split_1_10_shot: 58.9 vs 59.6(from the paper) voc_split_2_10_shot: 44.9 vs 46.6(from the paper)

hzhupku commented 2 years ago

Thank you for the reminder. I have updated the code.

hzhupku commented 2 years ago

For the performance in the paper, we report the results averaged over multiple random runs.