Awesome

Instance Segmentation Using Mask-RCNN

This reprository demonstrates training a Mask-RCNN network to perform instance segmentation and running an inference on a few test images. The network is trained on two classes - 'Person' and 'Car' using the COCO 2014 dataset.

Minimum Requirements

This repository has the following minimum requirements-

• MATLAB R2020b or later
• Deep Learning Toolbox
• Computer Vision Toolbox

3rd Party Requirements

Running the network training code from this repository, needs the following -

• COCO 2014 datset (images and annotations)
• COCO API

Overview

Instance segmentation is computer vision technique which involves detection and localization of objects while simultaneously generating a segmentation map for each of the detected instances.

There are several deep learning algorithms for instance segmentation, the most popular being Mask-RCNN. The Mask-RCNN network belongs to RCNN family of networks and builds on the Faster-RCNN network to perform pixel level segmentation on the detected objects. The Mask-RCNN network uses a Faster-RCNN network with -

• A more accurate sub-pixel level ROI pooling layer - ROIAlign
• A Mask branch for pixel level object segmentation. <img src="images/maskRCNN.png" width=800>

Perform Instance Segmentation using Mask-RCNN

Setup

1. Add path to the source directory -

addpath('./src/);

2. Download and load the pre-trained network -

datadir = tempdir;
url = 'https://www.mathworks.com/supportfiles/vision/data/maskrcnn_pretrained_person_car.mat';

helper.downloadTrainedMaskRCNN(url, datadir);
pretrained = load(fullfile(datadir, 'maskrcnn_pretrained_person_car.mat'));
net = pretrained.net;

% Extract Mask segmentation sub-network
maskSubnet = helper.extractMaskNetwork(net);

Predict

Perform prediction using the detectMaskRCNN function-

img = imread('visionteam.jpg');
executionEnvironment = "gpu";

[boxes, scores, labels, masks] = detectMaskRCNN(net, maskSubnet, img, params, executionEnvironment);

% Visualize results
overlayedImage = insertObjectMask(img, masks);
figure, imshow(overlayedImage)
showShape("rectangle", gather(boxes), "Label", labels, "LineColor",'r')

Training

To train your Mask-RCNN network, follow the steps outlined in the following examples.

• MaskRCNNTrainingExample.mlx - Example showing how to train the Mask-RCNN network.
• MaskRCNNParallelTrainingExample.mlx - Example showing how to train the Mask-RCNN network on multiple GPUs.

Helpers

This repository provides the following files to help create and train Mask-RCNN networks-

• src/CreateMaskRCNNConfig - Function to create a configuration structure for Mask-RCNN.
• src/CreateMaskRCNN - Function to create a Mask-RCNN netwwork with a resnet-101 backbone.
• detectMaskRCNN - Function to perform inference on Mask-RCNN network.

References

K. He, G. Gkioxari, P. Dolla ́r, and R. Girshick. Mask R-CNN. In ICCV, 2017.

Note

• "This network was trained using the COCO dataset, which was collected by the COCO Consortium (cocodataset.org)."