Few-shot YOLOv3: Few-shot Object Detection YOLOv3

网络修改by xinyuuliu 交流请联系作者qq522414928

环境 Python3.6 pytorch1.5+ easydict

Model base on it

由Few-shot Object Detection via Feature Reweighting 改进

使用数据VOC 20类 15个base类和5个few-shot类

Base Training

Modify Config

Change the cfg/fewyolov3_voc.data file

metayolo=1
metain_type=2
data=voc
neg = 1
rand = 0
novel = data/voc_novels.txt
novelid = 0
steps=-1,64000
scales=1,0.1
learning_rate = 0.001
meta = data/voc_traindict_full.txt
train = /data/liuxy67/Few-shot-Object-Detection/scripts/train.txt
valid = /data/liuxy67/Few-shot-Object-Detection/scripts/2007_test.txt
backup = backup/metayolov3_voc
gpus=0
num_workers=4

Train the Model

python train.py cfg/fewyolov3_voc.data cfg/darknet_yolov3_spp.cfg cfg/reweighting_net.cfg

Evaluate the Model

python valid.py cfg/fewyolov3_voc.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=voc
tuning = 1
neg = 0
rand = 0
novel = data/voc_novels.txt
novelid = 0
max_epoch = 2000
repeat = 200
dynamic = 0
scale=1
train = ./scripts/train.txt
meta = data/voc_traindict_bbox_10shot.txt
valid = ./scripts/2007_test.txt
backup = backup/metatune
gpus  = 0

Train the Model with 10 shot

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