Our experiments are based on the mmfewshot framework, plesase follow its instruction to install mmfewshot and mmdetection.
Our released dataset can be download from this link. It consists of 123287 npy files with a total size of 391 GB.
To integrate our dataset with mmfewshot, please use our functions (mmdet/datasets/pipelines) to load the dataset.
Our configurations for base training and few-shot learning by using Faster RCNN model are provided in configs/detection/tfa/coco.
Our test dataset is PASCALRAW, we apply the few-shot learning on the dataset containing the 10X downscaled images. You need to generate a class-blanced subset of the dataset as the training dataset, and use the rest as the test dataset.
Note, since we applied the demosaic function to the raw dataset, we need to divide the bbox by 2 in the annotation files for few-shot learning.