The goal is to train a model based on Matterport (Mask RCNN) to detect and segment an object. This is a Transfer learningis and based on COCO dataset. We should create our own dataset and annotate the object with Via annotator polygon. The object that I choose should not be included in the COCO dataset.I used Acoustic Guitar as a new object that is not in COCO dataset. I change some hyper parameters like the number of epochs and the learning rate to see the effect on the training results. Finally, the result of training should be declare as validation accuracy, training loss and some images that shows the segmentation and detection. In addition there is a part of discussion that is regarding to difficulties and the steps for solving them.
At the first step, the object for training and validation should be annotated by a human. I use https://www.robots.ox.ac.uk/~vgg/software/via/via_demo.html and Polygon selector to segment the region of Gitar. The output of these annotation is a Json file. ( you can find the images and Json file in the dataset (gitar ) part ). In addition, I used the Mask RCNN sample https://github.com/matterport/Mask_RCNN as a basic code but I cahnged some parts according to what I need.
I used Google Colab to run this project. The most important part of running the project is to check the address for Mask RCNN, dataset (gitar) images, and the JSON file in the main code. You need to download Mask RCNN project https://github.com/matterport/Mask_RCNN and upload it in your drive. Also, I upload dataset (gitar) and also Json files ( check the directory, for me Json files are in the Mask RCNN folder ) on my Google drive then I use them in my code. ( for example '/content/drive/MyDrive/Mask_RCNN' ). I did not use enough images for the training part (24 annotated images ). If you add more images you can get better results.
