bravercole/orefsdet

OreFSDet

OreFSDet is based on FewX ( an open source toolbox on top of Detectron2 for data-limited instance-level recognition tasks, e.g.)

OreFsdet and baseline on ore dataset

Method	5-shot		15-shot		25-shot
	AP	AP75	AP	AP75	AP	AP75
Attentionrpn(baseline)	25.1	27.0	29.2	34.5	30.8	37.0
orefsdet	36.2	33.0	39.3	45.6	44.7	48.4

The model can be obtained from here model .

Step 1: Installation

You only need to install detectron2. We recommend the Pre-Built Detectron2 (Linux only) version with pytorch 1.7. I use the Pre-Built Detectron2 with CUDA 10.1 and pytorch 1.7 and you can run this code to install it.

python -m pip install detectron2 -f \
  https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.7/index.html

Step 2: Prepare dataset

• Prepare for ore dataset, you can get from here. The ore dataset has been handled under few-shot setting, you only need to add it to dataset.

Step 3: Training and Evaluation

Run sh all.sh in the root dir.

Training

change all.sh

rm support_dir/support_feature.pkl
CUDA_VISIBLE_DEVICES=0 python3 fsod_train_net.py --num-gpus 1 \
	--config-file configs/fsod/finetune_R_50_C4_1x.yaml 2>&1 | tee log/fsod_finetune_stone_R50_train_log_5shot.txt

Then, run the following

sh all.sh

Evaluation

change the all.sh

CUDA_VISIBLE_DEVICES=0 python3 fsod_train_net.py --num-gpus 1 \
	--config-file configs/fsod/finetune_R_50_C4_1x.yaml \
	--eval-only MODEL.WEIGHTS ./output/fsod/finetune_dir/R_50_C4_1x_stone_5shot/model_final.pth 2>&1 | tee log/fsod_finetune_stone_R50_test_log_5shot.txt

just run the following

sh all.sh

Visualize the results

python demo.py  \
    --config-file configs/fsod/finetune_R_50_C4_1x.yaml \
    --input directory/*.png \
    --output results \
    --opts MODEL.WEIGHTS ./output/fsod/finetune_dir/R_50_C4_1x/model_final.pth

This repo is developed based on FewX and detectron2. Thanks for their wonderful codebases.