SIAnalytics / simplified_rbox_cnn

RBox-CNN

This is implementation of simplified version of RBox-CNN "RBox-CNN: rotated bounding box based CNN for ship detection in remote sensing image".
This code is based on Tensorflow Object Detection API and specially modified for the Object Detection Challenge in Satellite Image

Requirements

• python 3.5+
• tensorflow 1.6
• opencv-python 3.4
• protobuf 3.0

Installation

Protobuf Compliation(ref: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md)

protoc protos/*.proto --python_out=.

Dataset Preparation

We only provide a code to create tfrecords from GeoJson.
Data format:

   ├── Root
       ├── labels.geojson    
       ├── images  
           ├── *.png    

TFRecords for training are created from GeoJson by create_dataset.py

python create_dataset.py --src_dir='Root directory to geojson and images'
                         --dst_path='Path to save tfrecords'
                         ...

Training

Configuring the model pipeline

The config file of simplified RBox-CNN is "/configs/rbox_cnn_resnent101.config".
You have to set input path of tfrecords in the config file.
We recommend to set 'fine_tune_checkpoint_path' to the pretrained model(faster_rcnn_resnet101_coco_11_06_2017)
More parameters for training refer to Tensorflow Object Detection API

tf_record_input_reader {
  input_path: "/usr/home/username/data/train.tfrecord"
}
label_map_path: "/usr/home/username/data/label_map.pbtxt"

Running the training model

python train.py --pipeline_config_path='Path to a config file of the model'
                --train_dir='Directory to save the checkpoints and training summaries'
                ...
                

Evaluation

You can evaluate a trained model using tfrecords. The results of mAP and images are plotted by TensorBoard.
You have to set input path of tfrecords in the config file. More parameters for training refer to Tensorflow Object Detection API

eval_config:
{
    num_examples: Number of samples 
    eval_input_reader {
        input_path: "/usr/home/username/data/val.tfrecord"
    }
    label_map_path: "/usr/home/username/data/label_map.pbtxt"

Running the evaluating model

You can select a evaluation mode.

• 'latest': it run infinite and the latest checkpoint is evaluated.
• 'latest_once': latest checkpoint is evaluated and it finishes evaluation.
• 'all', all checkpoints are evaluated and it finishes evaluation.

python eval.py --pipeline_config_path='Path to a config file of the model'
                --checkpoint_dir='Directory to save the checkpoints and training summaries'
                --eval_dir='Directory to write eval summaries to'
                --run_mode='One of [latest, latest_once, all]'
                ...               

We provide a shell script('train_n_eval.sh') for train and evaluation.

Extras

We provide a submission code for Object Detection Challenge in Satellite Image. The submission code is "inference.py and run steps below.

1. Split images into small patch images
2. Infer patch images
3. Merge detected objects of patch images
4. Apply NMS to remove overlapping detected objects
5. Save CSV file for submission

python inference.py --pipeline_config_path='Path to a config file of the model'
                    --ckpt_path='Path to trained checkpoint, typically of the form path/to/model-%step.ckpt''
                    --image_dir='Path to images to be inferred'
                    --dst_path='Path to save detection output(CSV file)'
                    ...