search

lovelyqian / CDFSOD-benchmark

A benchmark for cross-domain few-shot object detection (ECCV24 paper: Cross-Domain Few-Shot Object Detection via Enhanced Open-Set Object Detector)
Apache License 2.0
28 stars 1 forks source link

Finetuning using the entire train dataset #7

Open 44526653 opened 1 month ago

44526653 commented 1 month ago

Hello. Thanks for the great work.

Instead of finetuning with k-shot, I want to finetune with the entire DIOR dataset(DIOR_train). What should I do? First, I tried to create a .pkl file for DIOR_train using build_prototypes.sh. However, I’m not sure how much —num_prototypes should be set. Also, after running build_prototypes.sh, a total of 3 pkl files are created. I would appreciate it if you could also tell me which pkl file should be used to set the class_prototypes of the .yaml file.

AImind commented 1 month ago

Hello! Thank you for your interest in our work.

devit use a clustering algorithm when generating prototypes. If you use more than 10shot data, you can also set num_prototypes to 10 in here. The file named ${dataset}_${shot}shot.vit${model}14.bbox.p${num_prototypes}.sk.pkl is the final prototypes file.

You can also refer to devit to adjust the strategy for generating prototypes.