stenlyho/mask-r-cnn-for-water-detection

Mask R-CNN

Mask-R-CNN-for-water-detection

This is an implimentation of Mask R-CNN proposed by Facebook AI Research (Girshick et al) for water detection implemented in python3, Keras and Tensorflow based on (https://arxiv.org/abs/1703.06870). The Mask R-CNN model performes water curtain detection, draws bounding box around it and applies segmentation mask.

The repository includes:

• The source code of Mask R-CNN with Resnet 101 and FPN(Feature Pyramid Network) as backbone.
• The implimentation detail of Mask R-CNN for water detection.
• A step by step procedure to train Mask R-CNN on your own dataset.
• Pre-trained weights of MS coco dataset.
• A Python code to run object detection on video and webcam.

The implimentation can be devided in to three steps.

1. Preparation of Datasets
2. Mask R-CNN Training
3. Testing the Model

1. Preparation of dataset

Mask R-CNN model training requires annotated images. This work uses VGG-Image annotator-1.0.6 for annotation, which can be downloaded from here. The VGG annotation tool is an online html web page which does not require any downloading and is very easy to use. The annotation tool gives annotation in .json file containg the x and y co-ordinates of annotated pixels.

CODE TIP:
If you are using VGG-Image annotator-1.0.6, than the code of this repository doesnot require any modification
while training.
Side Note: If your data set is in video instead of images, this repository includes the python code to sample
video at predefined sampling rate into images.