Clone and modified from https://github.com/yangxue0827/FPN_Tensorflow

Deleted some huge files from git source to keep it clean and easy to use.

Train a model using your own custom dataset

1.Place your dataset files in data/ folders, named layer for example

1.1 Place original images in data/layer/JPEGImages folder.

1.2 Place original annotations in data/layer/annotations folder.

1.3 Run gen_classes.py to generate classes file named classes.txt in data/layer/ folder.

1.3 Run convert_txt2xml.py to convert txt annotations to xml ones from folder data/layer/annotations to folder data/layer/Annotations.

1.4 Run data/io/convert_data_to_tfrecord.py to convert images and annotations to tfrecords files which located in folder data/tfrecords.

• run tools/test.py to label ground truth annotations on images to check if the data are right.

2.Modify configures

2.1 Modify libs/configs/cfgs.py to coordinate your own dataset: layer.

• Parameters

NET_NAME = 'resnet_v1_101'
DATASET_NAME = 'layer'
VERSION = 'v1_{}'.format(DATASET_NAME)
ANCHOR_SCALES = [0.5, 1., 2.]
ANCHOR_RATIOS = [0.1, 0.2, 0.3] # height to width
SCALE_FACTORS = [10., 5., 1., 0.5]

• Classes

CLASS_NUM = 1 #Equal to really class number (except for background class)

2.2 Modify configure file in folder configs according to NET_NAME

• configs/config_res101.py

pretrained_model_path # to use a pretrained model
batch_size

2.3 Add dataset name to data/io/read_tfrecord.py

• line 1 of function next_batch

['nwpu', 'airplane', 'SSDD', 'ship', 'pascal', 'coco', 'icecream', 'layer']

2.4 Add your NAME_LABEL_MAP corresponding to your own dataset in libs/label_name_dict/label_dict.py

• Directly add them if the number of classes is not big
• Or add NAME_LABEL_MAP using generated data/layer/classes.txt file.
• Examples:

elif cfgs.DATASET_NAME == 'icecream':
  NAME_LABEL_MAP = {}
  NAME_LABEL_MAP['back_ground'] = 0
  with open('classes.txt') as f:
    lines = [line.strip() for line in f.readlines()]
  for i, line in enumerate(lines, 1):
    NAME_LABEL_MAP[line] = i
elif cfgs.DATASET_NAME == 'layer':
  NAME_LABEL_MAP = {
      'back_ground': 0,
      "layer": 1
  }

3.Run scripts/train.sh to train the model and the output and logs will be saved in the root directory

cd $ FPN_Tensorflow
# ./scripts/train.sh GPU DATASET
./scripts/train.sh 0 cooler

4.Run scripts/[test.sh, eval.sh, demo.sh, inference.sh] to test, evaluate the model or run a demo using the trained model

cd $ FPN_Tensorflow
# ./scripts/test.sh GPU MODEL_PATH IMG_NUM
./scripts/test.sh 0 output/res101_trained_weights/v1_layer/layer_model.ckpt 20
# ./scripts/eval.sh GPU MODEL_PATH IMG_NUM
./scripts/eval.sh 0 output/res101_trained_weights/v1_layer/layer_model.ckpt 20
# ./scripts/demo.sh GPU MODEL_PATH
./scripts/demo.sh 0 output/res101_trained_weights/v1_layer/layer_model.ckpt
# ./scripts/inference.sh GPU MODEL_PATH
./scripts/inference.sh 0 output/res101_trained_weights/v1_layer/layer_model.ckpt

Errors may encountered

1.InvalidArgumentError (see above for traceback): LossTensor is inf or nan : Tensor had NaN values

InvalidArgumentError (see above for traceback): LossTensor is inf or nan : Tensor had NaN values
	 [[Node: train_op/CheckNumerics = CheckNumerics[T=DT_FLOAT, message="LossTensor is inf or nan", _device="/job:localhost/replica:0/task:0/device:GPU:0"](control_dependency)]]
	 [[Node: gradients/rpn_net/concat_grad/Squeeze_3/_1493 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_8085_gradients/rpn_net/concat_grad/Squeeze_3", tensor_type=DT_INT64, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

• This was raised because the annotations were beyond the images, for example, xmax or ymax larger than width or height of image, or xmin or ymin less than 0.
• This error has been solved by adding these lines in data/io/convert_data_to_tfrecord.py:

xmin = np.where(xmin < 0, 0, xmin)
ymin = np.where(ymin < 0, 0, ymin)
xmax = np.where(xmax > img_width, img_width, xmax)
ymax = np.where(ymax > img_height, img_height, ymax)

2.tensorflow.python.framework.errors_impl.UnknownError: exceptions.OverflowError: signed integer is less than minimum

UnknownError (see above for traceback): exceptions.OverflowError: signed integer is less than minimum
	 [[Node: fast_rcnn_loss/PyFunc_1 = PyFunc[Tin=[DT_FLOAT, DT_FLOAT, DT_INT32], Tout=[DT_UINT8], token="pyfunc_7", _device="/job:localhost/replica:0/task:0/device:CPU:0"](rpn_losses/Squeeze/_1579, fast_rcnn_loss/mul_1/_1759, fast_rcnn_loss/strided_slice_1/_1761)]]
	 [[Node: draw_proposals/Reshape_2/tensor/_1825 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_3802_draw_proposals/Reshape_2/tensor", tensor_type=DT_UINT8, _device="/job:localhost/replica:0/task:0/device:GPU:0"]()]]

• The reason of this error is the same as 1.InvalidArgumentError.

Feature Pyramid Networks for Object Detection

A Tensorflow implementation of FPN detection framework. You can refer to the paper Feature Pyramid Networks for Object Detection Rotation detection method baesd on FPN reference R2CNN and R2CNN_HEAD and R-DFPN If useful to you, please star to support my work. Thanks.

Configuration Environment

ubuntu(Encoding problems may occur on windows) + python2 + tensorflow1.2 + cv2 + cuda8.0 + GeForce GTX 1080 You can also use docker environment, command: docker push yangxue2docker/tensorflow3_gpu_cv2_sshd:v1.0

Installation

Clone the repository

git clone https://github.com/yangxue0827/FPN_Tensorflow.git

Make tfrecord

The image name is best in English. The data is VOC format, reference here data path format ($FPN_ROOT/data/io/divide_data.py) VOCdevkit

VOCdevkit_train

Annotation JPEGImages

VOCdevkit_test

Annotation JPEGImages

cd $FPN_ROOT/data/io/
python convert_data_to_tfrecord.py --VOC_dir='***/VOCdevkit/VOCdevkit_train/' --save_name='train' --img_format='.jpg' --dataset='ship'

Demo

1、Unzip the weight $FPN_ROOT/output/res101_trained_weights/*.rar 2、put images in $FPN_ROOT/tools/inference_image 3、Configure parameters in $FPN_ROOT/libs/configs/cfgs.py and modify the project's root directory 4、image slice

cd $FPN_ROOT/tools
python inference.py

5、big image

cd $FPN_ROOT/tools
python demo.py --src_folder=.\demo_src --des_folder=.\demo_des

Train

1、Modify $FPN_ROOT/libs/lable_name_dict/***_dict.py, corresponding to the number of categories in the configuration file 2、download pretrain weight(resnet_v1_101_2016_08_28.tar.gz or resnet_v1_50_2016_08_28.tar.gz) from here, then extract to folder $FPN_ROOT/data/pretrained_weights 3、

cd $FPN_ROOT/tools
python train.py

Test tfrecord

cd $FPN_ROOT/tools
python $FPN_ROOT/tools/test.py

eval

cd $FPN_ROOT/tools
python ship_eval.py

Summary

tensorboard --logdir=$FPN_ROOT/output/res101_summary/

Graph

Test results

airplane

sar_ship

ship

Note

This code works better when detecting single targets, but not suitable for multi-target detection tasks. Hope you can help find bugs, thank you very much.