search

ZhangGongjie / Meta-DETR

[T-PAMI 2022] Meta-DETR for Few-Shot Object Detection: Official PyTorch Implementation
MIT License
382 stars 83 forks source link

The implementation of instance segmentation. #47

Closed hq-deng closed 1 year ago

hq-deng commented 1 year ago

Could you please provide the code and checkpoints of few-shot instance segmentation? Thanks.

ZhangGongjie commented 1 year ago

Thanks for your interest. The codes I wrote for few-shot instance segmentation is a mess and are only used for a simple experiment.... Thus I do not have plans to release them.

hq-deng commented 1 year ago

Could your please send me a script for reference? No need to organize and verify. My email is "hanqiu1@ualberta.ca". Thanks.

ZhangGongjie commented 1 year ago

Done.

nanfangAlan commented 1 year ago

@ZhangGongjie Could you please send me the instance segmentation code also? My email is "nanfangalan@gmail.com". Thanks

ZhangGongjie commented 1 year ago

@nanfangAlan I have left Nanyang Technological University and thus I no longer have access to the codes. So sorry about that.

You may try to contact hanqiu1@ualberta.ca for the `messy' codes.

Besides, the implementation is actually very simple. It is the exact implementation as the original DETR, except that we designed the pipeline as a meta-learning setup. When the few-shot finetuning is done, we freeze all object-detection-related parameters and the backbone parameters and train an additional instance segmentation head.

Hope the above clarifies. Feel free to contact me for more details.

nanfangAlan commented 1 year ago

@ZhangGongjie Thanks for your reply. I'm checking the seg code in the original DETR. There are two more questions: 1. Which random seed sample can get the highest score and the "Results over a single run" score in your paper? 2. Did the other methods of "Results averaged over multiple random runs" take the same 10 random support sets as yours?