Boomerang-Watch Semantic Segmentation

We did this project for a masters course project at NUS - ISS (https://www.iss.nus.edu.sg/).

here we are showing the use of Mask RCNN in a real application.

We train the Mask RCNN model to detect Boomerang and Watch , and then we use the generated masks to predict the new images.

BW-Boomerang Watch Guitar is a defined but not trained class. Augment/change classes to your needs.

1. Installation

From the Releases page page:

1. Download mask_rcnn_coco.h5. Save it in the root directory of the repo (the mask_rcnn directory).
2. Download BW images. We used combination of Flickr, and Bing API for identifying images
3. We use VGG Image anotation tool and polygon masks to be created. Online tool: https://www.robots.ox.ac.uk/~vgg/software/via/via_demo.html
4. Put that in the path mask_rcnn\samples\train for train, 'mask_rcnn/samples/validate' for validate and 'mask_rcnn/samples/test' for test.

2. Train the MR-CNN model with pretrained weights

Train a new model starting from pre-trained COCO weights

python main.py train --dataset="./mask_rcnn/samples/train" --weights=coco

MRCNN_Boomerang_Watch_Semantic_Trainer.py: is to import necessary library/classes/functions from MR-CNN @ https://github.com/matterport/Mask_RCNN/releases and then train the network for BW using annotated image training dataset ".\mask_rcnn\samples\train"

Run Jupyter notebooks

MRCNN-Boomerang-Watch-Validate-Test.ipynb is used to validate, test and evaluate the trained model e.g. we had stored model at ".\mask_rcnn\samplesajay-bw\bw20200530T1137\mask_rcnn_bw_0020.h5