The official implementation of Novel Instances Mining with Pseudo-Margin Evaluation for Few-Shot Object Detection
The code is built on TFA
Requirements
- Linux with Python >= 3.6
- PyTorch >= 1.4
- torchvision that matches the PyTorch installation
- CUDA 10.0, 10.1, 10.2
- GCC >= 4.9
we provide the pre-trainied 10-shot final detecor
python3 -m tools.test_net --num-gpus 1 \
--config-file configs/PascalVOC-detection/split1/faster_rcnn_R_101_FPN_ft_all1_10shot.yaml \
--eval-only
To train a base detector, run
python3 -m tools.train_net --num-gpus 1 \
--config-file configs/PascalVOC-detection/split1/faster_rcnn_R_101_FPN_base1.yaml
fine-tune the detector with novel set
python3 -m tools.ckpt_surgery \
--src1 checkpoints/coco/faster_rcnn/faster_rcnn_R_101_FPN_base/model_final.pth \
--method randinit \
--save-dir checkpoints/coco/faster_rcnn/faster_rcnn_R_101_FPN_all
--coco
python3 -m tools.train_net --num-gpus 1 \
--config-file configs/COCO-detection/faster_rcnn_R_101_FPN_ft_all_1shot.yaml \
--opts MODEL.WEIGHTS $WEIGHTS_PATH
fine-tune the detector with pseudo set
python3 -m tools.genarate_pseudo --num-gpus 1
python3 -m tools.train_feature --num-gpus 1