Getting Started with Mask R-CNN for Instance Segmentation

Instance segmentation is an enhanced type of object detection that generates a segmentation map for each detected instance of an object. Instance segmentation treats individual objects as distinct entities, regardless of the class of the objects. In contrast, semantic segmentation considers all objects of the same class as belonging to a single entity.

Several deep learning algorithms exist to perform instance segmentation. One popular algorithm is Mask R-CNN, which expands on the Faster R-CNN network to perform pixel-level segmentation on the detected objects [1]. The Mask R-CNN algorithm can accommodate multiple classes and overlapping objects.

For an example that shows how to train a Mask R-CNN using Computer Vision Toolbox™, see Multiclass Instance Segmentation using Mask R-CNN.

Training Data

To train a Mask R-CNN, you need the following data.

DataDescription
RGB image

RGB images that serve as network inputs, specified as H-by-W-by-3 numeric arrays.

For example, this sample RGB image is a modified image from the CamVid data set [2] that has been edited to remove personally identifiable information.

RGB image of a street scene with vehicles and pedestrians

Ground-truth bounding boxes

Bounding boxes for objects in the RGB images, specified as a NumObjects-by-4 matrix, with rows in the format [x y w h]).

For example, the bboxes variable shows the bounding boxes of six objects in the sample RGB image.

bboxes =

   394   442    36   101
   436   457    32    88
   619   293   209   281
   460   441   210   234
   862   375   190   314
   816   271   235   305

Instance labels

Label of each instance, specified as a NumObjects-by-1 string vector or a NumObjects-by-1 cell array of character vectors.)

For example, the labels variable shows the labels of six objects in the sample RGB image.

labels =

  6×1 cell array

    {'Person' }
    {'Person' }
    {'Vehicle'}
    {'Vehicle'}
    {'Vehicle'}
    {'Vehicle'}

Instance masks

Masks for instances of objects. Mask data comes in two formats:

  • Binary masks, specified as a logical array of size H-by-W-by-NumObjects. Each mask is the segmentation of one instance in the image.

  • Polygon coordinates, specified as a NumObjects-by-2 cell array. Each row of the array contains the (x,y) coordinates of a polygon along the boundary of one instance in the image.

    The Mask R-CNN network requires binary masks, not polygon coordinates. To convert polygon coordinates to binary masks, use the poly2mask function. The poly2mask function sets pixels that are inside the polygon to 1 and sets pixels outside the polygon to 0. This code shows how to convert polygon coordinates in the masks_polygon variable to binary masks of size h-by-w-by-numObjects.

    denseMasks = false([h,w,numObjects]);
for i = 1:numObjects
    denseMasks(:,:,i) = poly2mask(masks_polygon{i}(:,1),masks_polygon{i}(:,2),h,w);
end

For example, this montage shows the binary masks of six objects in the sample RGB image.

Six binary masks showing the segmentation of two pedestrians and four vehicles

Visualize Training Data

To display the instance masks over the image, use the insertObjectMask. You can specify a color map so that each instance appears in a different color. This sample code shows how display the instance masks in the masks variable over the RGB image in the im variable using the lines color map. 

imOverlay = insertObjectMask(im,masks,'Color',lines(numObjects));
imshow(imOverlay);

Each pedestrian and vehicle has a unique falsecolor hue over the RGB image

To show the bounding boxes with labels over the image, use the showShape function. This sample code shows how to show labeled rectangular shapes with bounding box size and position data in the bboxes variable and label data in the labels variable.

imshow(imOverlay)
showShape("rectangle",bboxes,"Label",labels,"Color","red");

Red rectangles labeled 'Pedestrian' and 'Vehicle' surround instances of each object

Preprocess Data

Format and Resize Data

Use a datastore to read data. The datastore must return data as a 1-by-4 cell array in the format {RGB images, bounding boxes, labels, masks}. The size of the images, bounding boxes, and masks must match the input size of the network. If you need to resize the data, then you can use the imresize to resize the RGB images and masks, and the bboxresize function to resize the bounding boxes.

For more information, see Datastores for Deep Learning (Deep Learning Toolbox).

Form Mini-Batches of Data

Training a Mask R-CNN network requires a custom training loop. To manage the mini-batching of observations in a custom training loop, create a minibatchqueue (Deep Learning Toolbox) object from the datastore. The minibatchqueue object casts data to a dlarray (Deep Learning Toolbox) object that enables auto differentiation in deep learning applications. If you have a supported GPU, then a minibatchqueue object also moves data to the GPU.

The next (Deep Learning Toolbox) function yields the next mini-batch of data from the minibatchqueue.

Mask R-CNN Network Architecture

The Mask R-CNN network consists of two stages. The first is a region proposal network (RPN), which predicts object proposal bounding boxes based on anchor boxes. The second stage is an R-CNN detector that refines these proposals, classifies them, and computes the pixel-level segmentation for these proposals.

The Mask R-CNN model builds on the Faster R-CNN model, which you can create using fasterRCNNLayers. Replace the ROI max pooling layer with an roiAlignLayer that provides more accurate sub-pixel level ROI pooling. The Mask R-CNN network also adds a mask branch for pixel level object segmentation. For more information about the Faster R-CNN network, see Getting Started with R-CNN, Fast R-CNN, and Faster R-CNN.

This diagram shows a modified Faster R-CNN network on the left and a mask branch on the right.

Train Mask R-CNN Network

Train the model in a custom training loop. For each iteration:

  • Read the data for current mini-batch using the next (Deep Learning Toolbox) function.

  • Evaluate the model gradients using the dlfeval (Deep Learning Toolbox) function and a custom helper function that calculates the gradients and overall loss for batches of training data.

  • Update the network learnable parameters using a function such as adamupdate (Deep Learning Toolbox) or sgdmupdate (Deep Learning Toolbox).

For an example that shows how to train a Mask R-CNN using Computer Vision Toolbox, see Multiclass Instance Segmentation using Mask R-CNN.

References

[1] He, Kaiming, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. “Mask R-CNN.” ArXiv:1703.06870 [Cs], January 24, 2018. https://arxiv.org/pdf/1703.06870.

[2] Brostow, Gabriel J., Julien Fauqueur, and Roberto Cipolla. “Semantic Object Classes in Video: A High-Definition Ground Truth Database.” Pattern Recognition Letters 30, no. 2 (January 2009): 88–97. https://doi.org/10.1016/j.patrec.2008.04.005.

See Also

Apps

Functions

Objects

Related Topics

External Websites

Computer Vision Toolbox Documentation

Support