比赛官网: http://rscup.bjxintong.com.cn/#/theme/2
使用RRPN模型,ResNeXt-101-FPN作为backbone,mAP达到0.336(单模,由于最后时间不够未加任何trick)。模型主要参考:https://github.com/mjq11302010044/RRPN_pytorch ,需要针对本项目进行一定的修改。
# first, make sure that your conda is setup properly with the right environment# for that, check that `which conda`, `which pip` and `which python` points to the# right path. From a clean conda env, this is what you need to do
conda create -n rrpn_pytorch python=3.6
conda activate rrpn_pytorch
# this installs the right pip and dependencies for the fresh python
conda install ipython
# maskrcnn_benchmark and coco api dependencies
pip install ninja yacs cython matplotlib tqdm opencv-python (可以先which pip确认下pip在当前环境)
# follow PyTorch installation in https://pytorch.org/get-started/locally
(注意,这里一定要写=1.0或1.1,否则默认装1.2版本,后面一堆问题....)
conda install -c pytorch pytorch-nightly=1.0 torchvision=0.2.1 cudatoolkit=10.0(cudatoolkit的版本可以用nvcc -V看一下cuda编译器的版本,这两个最好一致!另外注意如果是RTX等图灵架构系列,一定不要用9.0版本的cudatoolkit,必须重装10.0,否则后面执行train_net.py时会报错)
# install pycocotoolscd ~/github
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
python setup.py build_ext install
# install RRPN
git clone https://github.com/clw5180/remote_sensing_object_detection_2019.git
cd RRPN_pytorch
# the following will install the lib with# symbolic links, so that you can modify
# the files if you want and won't need to
# re-build it
python setup.py build develop(一定注意:后期若有修改maskrcnn-benchmark文件夹下的代码,一定要用这条命令重新编译!目的是同步build文件夹下的内容)
# 后面这几步是RRPN项目特有的
python rotation_setup.py install
mv build/lib.linux-x86_64-3.6/rotation/*.so ./rotation (如果python版本不是3.6,就改成版本对应的数字)
# or if you are on macOS# MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py build develop
坑比较多,可以参考:
https://github.com/facebookresearch/maskrcnn-benchmark
https://zhuanlan.zhihu.com/p/64605565
采用800x800的crop,overlap取256,大约生成25000张图片。注意crop的同时要转换出对应的txt文件,放在labelTxt下。具体实现可以参考我的github,位置:Python_Practice/常用数据预处理脚本/超大图片crop/2019遥感_Step1_train_crop切图.py
1、在maskrcnn_benchmark/config/paths_catalog.py根据个人数据集的路径进行修改
"RRPN_train":
{
'dataset_list':
{
# 这里根据个人数据集的路径进行修改,该路径下含有放在labelTxt下和images两个文件,分别存放放标注数据和图片
# 这两个文件名可在在maskrcnn_benchmark/data/datasets/rotation_series.py中修改,见最后
'DOTA': '/media/clwclw/Elements/deep_learning/competion/2019yaogan/train/train_crop_800/'
},
"split": 'train'
},
2、tools文件夹下,执行python train_net.py
在demo文件夹下,执行python RRPN_Demo.py
主要是增加了在线crop,对bbox进行坐标变换,再用nms去掉overlap的重复检测框。
1、invalid value encountered in greater overlaps[overlaps > 1.00000001] = 0.0 解决办法: (1)删除小于16x16(保险起见建议8x8)的bbox; (2)在代码中找到T.RandomRotation,注释掉。
im_info = {
'gt_classes': gt_classes,
'max_classes': max_classes,
'image': im_path,
'boxes': gt_boxes,
'gt_words': gt_words,
'flipped': False,
'gt_overlaps': overlaps,
'seg_areas': seg_areas,
'height': im.shape[0],
'width': im.shape[1],
'max_overlaps': max_overlaps,
'rotated': True
}
im_infos.append(im_info)
...
return im_infos
同时也会将im_infos转换成pkl格式文件,然后保存在tools/data_cache下,
f_save_pkl = open(cache_pkl, 'wb')
pickle.dump(im_infos, f_save_pkl)
f_save_pkl.close()
print("Save pickle done.")
以便下一次再训练的时候可以直接读入:
if os.path.isfile(cache_pkl):
return pickle.load(open(cache_pkl, 'rb'))
这里我首先删掉了50kB以上的txt及对应的images(因为比如small-vehicle这个类别,在一张800x800的切图中最多能达到2500+的个数,所以很容易就会OOM),然后过滤掉了训练集中小于8x8的bbox(会在训练中引起一些overlap的错误或loss=nan的现象,另外这么小的尺度,下采样几次之后在feature map基本就找不到了,因此也没必要保留),。
elif "RRPN_train" in name:
data_dir = DatasetCatalog.DATA_DIR
attrs = DatasetCatalog.DATASETS['RRPN_train']
args = dict(
database=attrs["dataset_list"],
)
return dict(
factory="RotationDataset",
args=args,
)
当然也可以将数据集转化为COCO官方提供的json格式,只需要在maskrcnn_benchmark/config/paths_catalog.py中增加如下内容:
"DOTA_train": {
"img_dir": '/media/clwclw/Elements/deep_learning/competion/2019yaogan/train/train_crop_800/images/',
"ann_file": '/media/clwclw/Elements/deep_learning/competion/2019yaogan/train/train_crop_800/train.json',
"split": 'train'
},
elif "DOTA_train" in name:
data_dir = DatasetCatalog.DATA_DIR
attrs = DatasetCatalog.DATASETS[name]
args = dict(
root=os.path.join(data_dir, attrs["img_dir"]),
ann_file=os.path.join(data_dir, attrs["ann_file"]),
)
return dict(
factory="COCODataset", # 通过官方API导入json格式数据
args=args,
)
1、配置文件e2e_rrpn_X_101_32x8d_FPN_1x_DOTA.yaml需要根据情况进行修改:
DATASETS:
TRAIN: ("RRPN_train", )
......
ROI_BOX_HEAD:
NUM_CLASSES: 19 # 18类物体 + 1背景
OUTPUT_DIR: './models/DOTA/'
2、tools/train_net.py中,需要根据情况修改成自己的.yaml配置文件,如default="../configs/rrpn/e2e_rrpn_X_101_32x8d_FPN_1x_DOTA.yaml"
3、修改maskrcnn_benchmark/config/paths_catalog.py中需要添加自己的数据集,如
"XXXX_train": {
"img_dir":'/home/xxx/2019yaogan/train_data/images/',
"ann_file":'/home/xxx/2019yaogan/train_data/Annotations/train.json',
"split": 'train'
},
4、在maskrcnn_benchmark/data/datasets/rotation_series.py中需要修改的地方
def get_DOTA(mode, dataset_dir):
DATASET_DIR = dataset_dir
print('clw:在get_DOTA中, dataset_dir = ', dataset_dir)
# 如果dataset_dir下存放标注数据和图片的文件夹名称不是images和labelTxt,可以在下面进行修改
img_dir = "/images/"
gt_dir = "/labelTxt/"
DATASET = {
'IC13':get_ICDAR2013,
'IC15':get_ICDAR2015_RRC_PICK_TRAIN,
'IC17mlt':get_ICDAR2017_mlt,
'LSVT':get_ICDAR_LSVT_full,
'ArT':get_ICDAR_ArT,
'ReCTs':get_ICDAR_ReCTs_full,
'DOTA':get_DOTA, # 增加自己的数据集名称及对应获取数据的方法
}
class RotationDataset(torch.utils.data.Dataset):
CLASSES = (
"__background__ ", #"background",
"roundabout",
"tennis-court",
"swimming-pool",
"storage-tank",
"soccer-ball-field",
"small-vehicle",
"ship",
"plane",
"large-vehicle",
"helicopter",
"harbor",
"ground-track-field",
"bridge",
"basketball-court",
"baseball-diamond",
"helipad",
"airport",
"container-crane" # 修改自己的类别
)
结果如下:
写得比较仓促,如有遗漏或遇到问题欢迎提出宝贵issue,我会及时修正。 欢迎star!