/G-FSDet

Primary LanguagePython

Generalized Few-Shot Object Detection in Remote Sensing Images

This is the code for "Generalized Few-Shot Object Detection in Remote Sensing Images"

This code is based on MMFewshot, you can see the mmfew for more detail about the instructions.

Two-stage training framework

Following the original implementation, it consists of 3 steps:

  • Step1: Base training

    • use all the images and annotations of base classes to train a base model.

  • Step2: Reshape the bbox head of base model:

    • create a new bbox head for all classes fine-tuning (base classes + novel classes) using provided script.
    • the weights of base class in new bbox head directly use the original one as initialization.
    • the weights of novel class in new bbox head use random initialization.

  • Step3: Few shot fine-tuning:

    • use the base model from step2 as model initialization and further fine tune the bbox head with few shot datasets.

An example of DIOR split1 10-shot setting with 2 gpus

# step1: base training for voc split1
bash ./tools/detection/dist_train.sh \
    configs/detection/ETF/dior/split1/tfa_r101_fpn_dior-split1_base-training.py 2

# step2: reshape the bbox head of base model for few shot fine-tuning
python -m tools.detection.misc.initialize_bbox_head \
    --src1 work_dirs/ETF_r101_fpn_voc-split1_base-training/latest.pth \
    --method randinit \
    --save-dir work_dirs/ETF_r101_fpn_voc-split1_base-training

# step3(Model ETF): few shot fine-tuning
bash ./tools/detection/dist_train.sh \
    configs/detection/ETF/dior/split1/ETF_r101_fpn_dior-split1_10shot-fine-tuning.py 2


# step3(Model ETF+Dis): few shot fine-tuning
bash ./tools/detection/dist_train.sh \
    configs/detection/dis_loss/dior/split1/power4_dis_tfa_r101_fpn_dior-split1_3shot-fine-tuning.py 2


# step3(Model G-FSDet): few shot fine-tuning
bash ./tools/detection/dist_train.sh \
    configs/detection/GFSDet/dior/split1/power4_0.025_weight_0.5_alpha_tfa_r101_fpn_dior-split1_3shot-fine-tuning.py 2

Note:

  • The default output path of the reshaped base model in step2 is set to work_dirs/{BASE TRAINING CONFIG}/base_model_random_init_bbox_head.pth. When the model is saved to different path, please update the argument load_from in step3 few shot fine-tune configs instead of using resume_from.
  • To use pre-trained checkpoint, please set the load_from to the downloaded checkpoint path.

Data preparation

We have provided the few-shot annotations in 'data/few_shot_ann'.

Base training checkpoint on DIOR dataset

Split1: https://pan.baidu.com/s/11PcX-ywOiF3bPhFIcZKlUA
code：ouhu