fsod

few shot object detectioin with GNN

Code built on top of MetaR-CNN.

requirements

Python = 3.6 CUDA = 8.0 PyTorch = 0.4.0 torchvision = 0.2.1 gcc >= 4.9

environment

conda create --name fsod --file spec-file.txt conda activate fsod

compile CUDA dependencies

cd {repo_root}/lib sh make.sh

data

Download data from from google drive and put custom folder in {repo_root}/data/ https://drive.google.com/drive/folders/1lQpiZRoOD6L81h_MfWHbjmGVt2K5spRu?usp=sharing

A trainval.csv file containing all the annotations such as: set,im_path,cls,difficult,left,upper,right,lower,height,width train,ImagePathInDataset,cat,False,x1,y1,x2,y2,H,W train,ImagePathInDataset,dog,False,x1,y1,x2,y2,H,W val,ImagePathInDataset,cat,False,x1,y1,x2,y2,H,W val,ImagePathInDataset,dog,False,x1,y1,x2,y2,H,W cls contains base and novel class: 'Brake', 'BuckleA', 'Fixator', 'JointB', 'PlateA', 'Spring', 'WheelA', 'PlateD', 'PlateB', 'WheelB', 'BuckleB', 'Bearing', 'PlateC'

If you want change categories of base class change in line 27 {repo_root}/lib/model/utils/config.py eg: __C.CUSTOM_NOVEL_CLASSES = ['PlateA','PlateB','BuckleA','Brake','WheelA'] change number of categories in line207 {repo_root}/train.py eg: base_num = 8

prepare pre-trained ResNet model

Download data from from google drive and put pretrained_model folder in {repo_root}/data/ https://drive.google.com/drive/folders/1aLahARQf1SH_NSm1gyd_0h5QTTMzBlEn?usp=sharing

Download prepare trained model from from google drive and put save_models folder in {repo_root}/

https://drive.google.com/drive/folders/10hZ_PUxlLl0_OEAGpB8TsPQOeWi9ffc2?usp=sharing

train

base train bash run/train_voc_custom.sh
fine-tune train bash run/finetune_voc_custom.sh

test

bash run/test_voc_custom.sh

result

You can find the result in {repo_root}/save_models/custom/custom/

ChengZY/fsod