This is the code of paper "Fine-Grained Recognition for Oriented Ship Against Complex Scenes in Optical Remote in Remote Sensing Images".
Efficient Information Reuse Network (EIRNet) is a two-stage end-to-end network designed for fine-grained ship recognition in remote sensing images.
We tested EIRNet on two datasets: HRSC2016 and DOSR. HRSC2016 is a widely used dataset for fine-grained ship recognition. DOSR is a new fine-grained ship recognition dataset collected by ourselves with more complex scenes. Extensive experiments demonstrate our EIRNet achieves state-of-the-art performance on DOSR and HRSC2016.
Ubuntu / Win10
Python 3.5
Tensorflow 1.10.0
Cuda 9.0
GeForce GTX 1080Ti
git clone https://github.com/yaqihan-9898/EIRNet # clone repo
cd EIRNet
pip install -r requirements.txt # install
For linux:
cd ./data/coco/PythonAPI
python setup.py build_ext --inplace
python setup.py build_ext install
cd ./lib/utils
python setup.py build_ext --inplace
For windows:
cd /d .\data\coco\PythonAPI
python setup.py build_ext --inplace
python setup.py build_ext install
cd /d .\lib\utils
python setup.py build_ext --inplace
Put all dataset images under ./data/VOCdevkit2007/VOC2007/JPEGImages
Put all dataset annotations (VOC format) under ./data/VOCdevkit2007/VOC2007/Annotations
Put train.txt, val.txt, test.txt, and trainval.txt under ./data/VOCdevkit2007/VOC2007/ImageSets/Main
Download pretrained weight from our pretrained model (based on ResNet101) with access code 1596 or office model.
Modify ./lib/config/config.py:
'network' in Line10
If you want to train your own dataset, please modify:
./lib/config/config.py: 'dataset' in Line20, 'image_ext' in Line21, 'uplevel_len' in Line22
./lib/datasets/pascal_voc.py: 'CLASSES' in Line24
./lib/layer_utils/uplevel_utils.py: 'uplevelmap' in Line4
./lib/layer_utils/cls_weight.py: 'cls_weight_L2' in Line4
The default cache files in ./data/cache and ./data/VOCdevkit2007/annotations_cache is obtained from DOSR. If you want to train or test other datasets, please delete ./data/cache/HRSC2016/voc_2007_trainval_gt_roidb.pkl and ./data/VOCdevkit2007/annotations_cache/annots.pkl
Go to the root path and start to train:
$ python train.py
Note: At the beginning of the train, there may be some assert errors. The reason is there is an assert 'fg_num>0' in function 'iou_rotate_calculate3' of ./lib/utils/iou_rotate.py. At the beginning of the train, the network gets no foreground，thus, network cannot comput IoU between the predicted foreground box and its corresponding groundtruth bounding box. Such assert errors will not affect training process, you can ignore it. You can use our pretrained model to reduce such assert errors.
python demo.py
python eval_r.py