This reposotiries uses the Keras_MaskRCNN by Fizyr. For more detail about the Keras MaskRCNN, you're suggested to follow the original repo. In this repo, we utilize the data available in Airbus Ship Detection Competition hosted in kaggle to detect and segment the ship's instances in the statlite images.

We don't exactly modify origianl Keras MaskRCNN implementation exactly, however we do prepare the data required to fit the model, and perform post-processing of the result and turn that into the result we would get from Matterport's MASKRCNN implementation.

1. Clone this repository.
2. Install keras-retinanet (pip install keras-retinanet --user). Make sure tensorflow is installed and is using the GPU.
3. Optionally, install pycocotools if you want to train / test on the MS COCO dataset by running pip install --user git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI.
4. Clone this repo.

1. Download the Kaggle DataSet From the Airbus Ship Detection Competition page.

2. Extract all of them inside the data folder. You should have two folder, train_v2 and test_v2 containing the training and testing image dataset. Also you will have two csv files, one is train_ship_segmentation_v2.csv and another is sample_submission_v2.csv.

3. Now it's turn to create annotation from Run length Encoding and the mask files associated with each rle. For this you would have to run python createdata.py. Please make sure the data path are correct by checking the createdata.py script.

4. We don't need the csv header, so please remove the header column names from the annotation.csv and class_ids.csv if there is any. For your convenience sample_annotation.csv and class_ids.csv is already inside the data folder. However make sure to remove the header column from the csv file as they are not needed by the CsvDataGenerator. You may as well make a validation_annotation.csv.

Since our training data is ready inside our data folder, it's time to train the model. For our convinience, we have added a config folder inside the keras_maskrcnn and placed a config.ini, where we have changed the anchor size, because ship instances were mostly small and we wanted to detect all the small ships.

Assuming we have the annotation.csv, class_ids.csv, we can now proceed to the training.

1. Download the Coco weight provided from the Fizyr release. We used the resnet_50_coco_v_0.2.0.h5
2. Run the following command inside the project directory: ./keras_maskrcnn/bin/train.py --weights=PATH/TO/COCO/WEIGHT/ --epochs=50 --steps=1000 --config=config/config.ini csv data/annotation.csv data/class_ids.csv

In the model output boxes are shaped (None, None, 4) (for (x1, y1, x2, y2)), scores is shaped (None, None) (classification score), labels is shaped (None, None) (label corresponding to the score) and masks is shaped (None, None, 28, 28). In all three outputs, the first dimension represents the shape and the second dimension indexes the list of detections. In our case, we modled the mask of the result into (NUM_DETECTION, IMAGE_HEIGHT, IMAGE_WIDTH, 1), here IMAGE_HEIGHT and IMAGE_WIDTH is 768 and 768. We applied the Non-Max Suppression and removed the overlapping in the instanc segmented. The detail of the process is in bin/detect.py .

Lots of references of codes and implemntation were taken from the various sources, which are not least but listed below:

1. [https://github.com/fizyr/keras-maskrcnn/ ]
2. [https://github.com/matterport/Mask_RCNN] Many of the utility functions were actually the modified version available in the utility.py of matterport's Mask RCNN.
3. [https://github.com/mirzaevinom/data_science_bowl_2018]. Some of the utility functions were actually modified/used from mirzaevinom implementation.
4. Kaggle Airbus Ship Detection competition, Kernels.