Instance Segmentation

Perform instance segmentation using pretrained deep learning networks and train networks using transfer learning on custom data

Instance segmentation is a computer vision technique that plays a crucial role in tasks requiring precise object localization and the identification of individual object instances, such as in medical imaging and autonomous driving. By combining the principles of object detection and semantic segmentation, instance segmentation provides a more refined understanding of visual data by identifying individual object instances and delineating their boundaries pixel by pixel. Use instance segmentation to precisely identify, classify, and separate individual objects within an image.

You can perform inference on an image using a pretrained deep learning network, or train a network using perform transfer learning. Transfer learning is an approach in which you to start with a pretrained network and train it on a custom data set for your application. You can interactively generate ground truth data for training,by using the Image Labeler, Video Labeler, or Ground Truth Labeler (Automated Driving Toolbox) app to label pixels and export label data. Instance segmentation requires Deep Learning Toolbox™. Training and inference support CUDA^® enabled GPU. Use of a GPU is recommended, and requires Parallel Computing Toolbox™. For more information, see Parallel Computing Support in MathWorks Products (Parallel Computing Toolbox).

Instance segmentation using SOLOv2: Left — segmented and labeled road scenario using a sample modified RGB image from the CamVid data set, Right — segmented image of PVC pipe connectors

Functions

expand all

Segment Objects in Images

Configure Instance Segmentation Network

`solov2`	Segment objects using SOLOv2 instance segmentation network (Since R2023b)
`maskrcnn`	Detect objects using Mask R-CNN instance segmentation (Since R2021b)

Perform Inference

`segmentObjects`	Segment objects using Mask R-CNN instance segmentation (Since R2021b)
`segmentObjects`	Segment objects using SOLOv2 instance segmentation (Since R2023b)

Train Custom Instance Segmentation Networks

Load Training Data

`boxLabelDatastore`	Datastore for bounding box label data (Since R2019b)
`groundTruth`	Ground truth label data
`imageDatastore`	Datastore for image data
`combine`	Combine data from multiple datastores

Train Instance Segmentation Networks

`trainSOLOV2`	Train SOLOv2 network to perform instance segmentation (Since R2023b)
`trainMaskRCNN`	Train Mask R-CNN network to perform instance segmentation (Since R2022a)

Augment and Preprocess Training Data

`poly2mask`	Convert region of interest (ROI) polygon to region mask
`bwboundaries`	Trace object boundaries in binary image
`balanceBoxLabels`	Balance bounding box labels for object detection (Since R2020a)
`bboxcrop`	Crop bounding boxes (Since R2019b)
`bboxerase`	Remove bounding boxes (Since R2021a)
`bboxresize`	Resize bounding boxes (Since R2019b)
`bboxwarp`	Apply geometric transformation to bounding boxes (Since R2019b)
`bbox2points`	Convert rectangle to corner points list
`imwarp`	Apply geometric transformation to image
`imcrop`	Crop image
`imresize`	Resize image
`randomAffine2d`	Create randomized 2-D affine transformation (Since R2019b)
`centerCropWindow2d`	Create rectangular center cropping window (Since R2019b)
`randomWindow2d`	Randomly select rectangular region in image (Since R2021a)

Visualize Results

`insertObjectMask`	Insert masks in image or video stream (Since R2020b)
`insertObjectAnnotation`	Annotate truecolor or grayscale image or video
`insertShape`	Insert shapes in image or video
`showShape`	Display shapes on image, video, or point cloud (Since R2020b)

Evaluate Predicted Results

`evaluateInstanceSegmentation`	Evaluate instance segmentation data set against ground truth (Since R2022b)
`instanceSegmentationMetrics`	Instance segmentation quality metrics (Since R2022b)
`metricsByArea`	Evaluate instance segmentation across object mask size ranges (Since R2023b)

Perform Pose Estimation Using Instance Segmentation

`posemaskrcnn`	Predict object pose using Pose Mask R-CNN pose estimation (Since R2024a)
`predictPose`	Estimate object pose using Pose Mask R-CNN deep learning network (Since R2024a)
`trainPoseMaskRCNN`	Train Pose Mask R-CNN network to perform pose estimation (Since R2024a)

Topics

Get Started

Get Started with Instance Segmentation Using Deep Learning
Segment objects using an instance segmentation model such as SOLOv2 or Mask R-CNN.
Get Started with SOLOv2 for Instance Segmentation
Perform multiclass instance segmentation using SOLOv2 and deep learning.
Getting Started with Mask R-CNN for Instance Segmentation
Perform multiclass instance segmentation using Mask R-CNN and deep learning.

Train Data for Instance Segmentation

Get Started with Image Preprocessing and Augmentation for Deep Learning
Preprocess data for deep learning applications with deterministic operations such as resizing, or augment training data with randomized operations such as random cropping.
Datastores for Deep Learning (Deep Learning Toolbox)
Learn how to use datastores in deep learning applications.

Featured Examples

Perform Instance Segmentation Using SOLOv2

Segment object instances of randomly rotated machine parts in a bin using a deep learning SOLOv2 network.

Open Live Script

Perform Instance Segmentation Using Mask R-CNN

Segment individual instances of people and cars using a multiclass mask region-based convolutional neural network (R-CNN).

Open Live Script

Perform 6-DoF Pose Estimation for Bin Picking Using Deep Learning

Perform six degrees-of-freedom (6-DoF) pose estimation by estimating the 3-D position and orientation of machine parts in a bin using RGB-D images and a deep learning network.

Open Live Script