Few-shot YOLOv3: Few-shot Object Detection on Optical Remote Sensing Images

Our code is based on https://github.com/marvis/pytorch-yolo2 and developed with Python 2.7 & PyTorch 0.3.1.

Detection Examples

Samples of novel class detection result on NWPU VHR-10 and DIOR with 10-shot training bounding boxes.

Model

Training our model on NWPU VHR-10

$PROJ_ROOT : project root
$DATA_ROOT : dataset root

Prepare dataset

Get the NWPU VHR-10 data from https://1drv.ms/u/s!AmgKYzARBl5cczaUNysmiFRH4eE

Preprocess data

cd $PROJ_ROOT
python scripts/ImgSplit_nwpu.py $DATA_ROOT

Generate labels for NWPU VHR-10

python scripts/label_nwpu.py $DATA_ROOT

Generate per-class labels for NWPU VHR-10 (used for the Reweighting Module input)

python scripts/label_1c_nwpu.py $DATA_ROOT

Generate few-shot datasets
Change the ''DROOT'' varibale in scripts/gen_fewlist_nwpu.py to $DATA_ROOT

python scripts/gen_fewlist_nwpu.py

Generate training dictionary

python scripts/gen_dict_file.py $DATA_ROOT nwpu

Base Training

Modify Config for NWPU VHR-10 Data
Change the cfg/fewyolov3_nwpu.data file

metayolo = 1
metain_type = 2
data = nwpu
neg = 1
rand = 0
novel = data/nwpu_novels.txt
novelid = 0
learning_rate = 0.001
steps = -1,64000
scales = 1,0.1
meta = data/nwpu_traindict_full.txt
train = $DATA_ROOT/training.txt
valid = $DATA_ROOT/evaluation.txt
backup = backup/fewyolov3_nwpu
gpus = 0,1,2,3

Train the Model

python train.py cfg/fewyolov3_nwpu.data cfg/darknet_yolov3_spp.cfg cfg/reweighting_net.cfg

Evaluate the Model

python valid.py cfg/fewyolov3_nwpu.data cfg/darknet_yolov3_spp.cfg cfg/reweighting_net.cfg path/toweightfile
python scripts/voc_eval.py results/path/to/comp4_det_test_ cfg/metayolo.data

Few-shot Tuning

Modify Config for NWPU VHR-10 Data
Change the cfg/fewtunev3_nwpu_10shot.data file (change the shot number to try different shots)

metayolo = 1
metain_type = 2
data = nwpu
tuning = 1
neg = 0
rand = 0
novel = data/nwpu_novels.txt
novelid = 0
max_epoch = 2000
repeat = 200
dynamic = 0
scale = 1
learning_rate = 0.0001
steps = -1,64000
scales = 1,0.1
train = $DATA_ROOT/training.txt
meta = data/nwpu_traindict_bbox_10shot.txt
valid = $DATA_ROOT/evaluation.txt
backup = backup/fewtunetestv3_nwpu_10shot
gpus = 0,1,2,3

Train the Model with 10 shot

python train.py cfg/fewtunev3_nwpu_10shot.data cfg/darknet_yolov3_spp.cfg cfg/reweighting_net.cfg path/to/base/weightfile

Evaluate the Model

python valid.py cfg/fewtunev3_nwpu_10shot.data cfg/darknet_yolov3_spp.cfg cfg/reweighting_net.cfg path/to/tuned/weightfile
python scripts/voc_eval.py results/path/to/comp4_det_test_ cfg/fewtunev3_nwpu_10shot.data

Acknowledgements

Large part of the code is borrowed from YOLO-Low-Shot

Citation

If you find this useful in your research, please consider citing:

@article{deng2020few, title={Few-shot Object Detection on Remote Sensing Images}, author={Deng, Jingyu and Li, Xiang and Fang, Yi}, journal={arXiv preprint arXiv:2006.07826}, year={2020} }

lixiang-ucas / FSODM