How to Train an Object Detection Model with Keras

Last Updated on September 2, 2020

Object detection is a challenging computer vision task that involves predicting both where the objects are in the image and what type of objects were detected.

The Mask Region-based Convolutional Neural Network, or Mask R-CNN, model is one of the state-of-the-art approaches for object recognition tasks. The Matterport Mask R-CNN project provides a library that allows you to develop and train Mask R-CNN Keras models for your own object detection tasks. Using the library can be tricky for beginners and requires the careful preparation of the dataset, although it allows fast training via transfer learning with top performing models trained on challenging object detection tasks, such as MS COCO.

In this tutorial, you will discover how to develop a Mask R-CNN model for kangaroo object detection in photographs.

After completing this tutorial, you will know:

• How to prepare an object detection dataset ready for modeling with an R-CNN.
• How to use transfer learning to train an object detection model on a new dataset.
• How to evaluate a fit Mask R-CNN model on a test dataset and make predictions on new photos.

Kick-start your project with my new book Deep Learning for Computer Vision, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

How to Train an Object Detection Model to Find Kangaroos in Photographs (R-CNN with Keras)
Photo by Ronnie Robertson, some rights reserved.

Tutorial Overview

This tutorial is divided into five parts; they are:

1. How to Install Mask R-CNN for Keras
2. How to Prepare a Dataset for Object Detection
3. How to a Train Mask R-CNN Model for Kangaroo Detection
4. How to Evaluate a Mask R-CNN Model
5. How to Detect Kangaroos in New Photos

Note: This tutorial requires TensorFlow version 1.15.3 and Keras 2.2.4. It does not work with TensorFlow 2.0+ or Keras 2.2.5+ because a third-party library has not been updated at the time of writing.

You can install these specific versions of the libraries as follows:

How to Install Mask R-CNN for Keras

Object detection is a task in computer vision that involves identifying the presence, location, and type of one or more objects in a given image.

It is a challenging problem that involves building upon methods for object recognition (e.g. where are they), object localization (e.g. what are their extent), and object classification (e.g. what are they).

The Region-Based Convolutional Neural Network, or R-CNN, is a family of convolutional neural network models designed for object detection, developed by Ross Girshick, et al. There are perhaps four main variations of the approach, resulting in the current pinnacle called Mask R-CNN. The Mask R-CNN introduced in the 2018 paper titled “Mask R-CNN” is the most recent variation of the family of models and supports both object detection and object segmentation. Object segmentation not only involves localizing objects in the image but also specifies a mask for the image, indicating exactly which pixels in the image belong to the object.

Mask R-CNN is a sophisticated model to implement, especially as compared to a simple or even state-of-the-art deep convolutional neural network model. Instead of developing an implementation of the R-CNN or Mask R-CNN model from scratch, we can use a reliable third-party implementation built on top of the Keras deep learning framework.

The best-of-breed third-party implementations of Mask R-CNN is the Mask R-CNN Project developed by Matterport. The project is open source released under a permissive license (e.g. MIT license) and the code has been widely used on a variety of projects and Kaggle competitions.

The first step is to install the library.

At the time of writing, there is no distributed version of the library, so we have to install it manually. The good news is that this is very easy.

Installation involves cloning the GitHub repository and running the installation script on your workstation. If you are having trouble, see the installation instructions buried in the library’s readme file.

Step 1. Clone the Mask R-CNN GitHub Repository

This is as simple as running the following command from your command line:

This will create a new local directory with the name Mask_RCNN that looks as follows:

Step 2. Install the Mask R-CNN Library

The library can be installed directly via pip.

Change directory into the Mask_RCNN directory and run the installation script.

From the command line, type the following:

On Linux or MacOS, you may need to install the software with sudo permissions; for example, you may see an error such as:

In that case, install the software with sudo:

If you are using a Python virtual environment (virtualenv), such as on an EC2 Deep Learning AMI instance (recommended for this tutorial), you can install Mask_RCNN into your environment as follows:

The library will then install directly and you will see a lot of successful installation messages ending with the following:

This confirms that you installed the library successfully and that you have the latest version, which at the time of writing is version 2.1.

Step 3: Confirm the Library Was Installed

It is always a good idea to confirm that the library was installed correctly.

You can confirm that the library was installed correctly by querying it via the pip command; for example:

You should see output informing you of the version and installation location; for example:

We are now ready to use the library.

How to Prepare a Dataset for Object Detection

Next, we need a dataset to model.

In this tutorial, we will use the kangaroo dataset, made available by Huynh Ngoc Anh (experiencor). The dataset is comprised of 183 photographs that contain kangaroos, and XML annotation files that provide bounding boxes for the kangaroos in each photograph.

The Mask R-CNN is designed to learn to predict both bounding boxes for objects as well as masks for those detected objects, and the kangaroo dataset does not provide masks. As such, we will use the dataset to learn a kangaroo object detection task, and ignore the masks and not focus on the image segmentation capabilities of the model.

There are a few steps required in order to prepare this dataset for modeling and we will work through each in turn in this section, including downloading the dataset, parsing the annotations file, developing a KangarooDataset object that can be used by the Mask_RCNN library, then testing the dataset object to confirm that we are loading images and annotations correctly.

Install Dataset

The first step is to download the dataset into your current working directory.

This can be achieved by cloning the GitHub repository directly, as follows:

This will create a new directory called “kangaroo” with a subdirectory called ‘images/‘ that contains all of the JPEG photos of kangaroos and a subdirectory called ‘annotes/‘ that contains all of the XML files that describe the locations of kangaroos in each photo.

Looking in each subdirectory, you can see that the photos and annotation files use a consistent naming convention, with filenames using a 5-digit zero-padded numbering system; for example:

This makes matching photographs and annotation files together very easy.

We can also see that the numbering system is not contiguous, that there are some photos missing, e.g. there is no ‘00007‘ JPG or XML.

This means that we should focus on loading the list of actual files in the directory rather than using a numbering system.

Parse Annotation File

The next step is to figure out how to load the annotation files.

First, open the first annotation file (annots/00001.xml) and take a look; you should see:

We can see that the annotation file contains a “size” element that describes the shape of the photograph, and one or more “object” elements that describe the bounding boxes for the kangaroo objects in the photograph.

The size and the bounding boxes are the minimum information that we require from each annotation file. We could write some careful XML parsing code to process these annotation files, and that would be a good idea for a production system. Instead, we will short-cut development and use XPath queries to directly extract the data that we need from each file, e.g. a //size query to extract the size element and a //object or a //bndbox query to extract the bounding box elements.

Python provides the ElementTree API that can be used to load and parse an XML file and we can use the find() and findall() functions to perform the XPath queries on a loaded document.

First, the annotation file must be loaded and parsed as an ElementTree object.

Once loaded, we can retrieve the root element of the document from which we can perform our XPath queries.

We can use the findall() function with a query for ‘.//bndbox‘ to find all ‘bndbox‘ elements, then enumerate each to extract the x and y, min and max values that define each bounding box.

The element text can also be parsed to integer values.

We can then collect the definition of each bounding box into a list.

The dimensions of the image may also be helpful, which can be queried directly.

We can tie all of this together into a function that will take the annotation filename as an argument, extract the bounding box and image dimension details, and return them for use.

The extract_boxes() function below implements this behavior.

We can test out this function on our annotation files, for example, on the first annotation file in the directory.

The complete example is listed below.

Running the example returns a list that contains the details of each bounding box in the annotation file, as well as two integers for the width and height of the photograph.

Now that we know how to load the annotation file, we can look at using this functionality to develop a Dataset object.

Develop KangarooDataset Object

The mask-rcnn library requires that train, validation, and test datasets be managed by a mrcnn.utils.Dataset object.

This means that a new class must be defined that extends the mrcnn.utils.Dataset class and defines a function to load the dataset, with any name you like such as load_dataset(), and override two functions, one for loading a mask called load_mask() and one for loading an image reference (path or URL) called image_reference().

To use a Dataset object, it is instantiated, then your custom load function must be called, then finally the built-in prepare() function is called.

For example, we will create a new class called KangarooDataset that will be used as follows:

The custom load function, e.g. load_dataset() is responsible for both defining the classes and for defining the images in the dataset.

Classes are defined by calling the built-in add_class() function and specifying the ‘source‘ (the name of the dataset), the ‘class_id‘ or integer for the class (e.g. 1 for the first lass as 0 is reserved for the background class), and the ‘class_name‘ (e.g. ‘kangaroo‘).

Objects are defined by a call to the built-in add_image() function and specifying the ‘source‘ (the name of the dataset), a unique ‘image_id‘ (e.g. the filename without the file extension like ‘00001‘), and the path for where the image can be loaded (e.g. ‘kangaroo/images/00001.jpg‘).

This will define an “image info” dictionary for the image that can be retrieved later via the index or order in which the image was added to the dataset. You can also specify other arguments that will be added to the image info dictionary, such as an ‘annotation‘ to define the annotation path.

For example, we can implement a load_dataset() function that takes the path to the dataset directory and loads all images in the dataset.

Note, testing revealed that there is an issue with image number ‘00090‘, so we will exclude it from the dataset.

We can go one step further and add one more argument to the function to define whether the Dataset instance is for training or test/validation. We have about 160 photos, so we can use about 20%, or the last 32 photos, as a test or validation dataset and the first 131, or 80%, as the training dataset.

This division can be made using the integer in the filename, where all photos before photo number 150 will be train and equal or after 150 used for test. The updated load_dataset() with support for train and test datasets is provided below.

Next, we need to define the load_mask() function for loading the mask for a given ‘image_id‘.

In this case, the ‘image_id‘ is the integer index for an image in the dataset, assigned based on the order that the image was added via a call to add_image() when loading the dataset. The function must return an array of one or more masks for the photo associated with the image_id, and the classes for each mask.

We don’t have masks, but we do have bounding boxes. We can load the bounding boxes for a given photo and return them as masks. The library will then infer bounding boxes from our “masks” which will be the same size.

First, we must load the annotation file for the image_id. This involves first retrieving the ‘image info‘ dict for the image_id, then retrieving the annotations path that we stored for the image via our prior call to add_image(). We can then use the path in our call to extract_boxes() developed in the previous section to get the list of bounding boxes and the dimensions of the image.

We can now define a mask for each bounding box, and an associated class.

A mask is a two-dimensional array with the same dimensions as the photograph with all zero values where the object isn’t and all one values where the object is in the photograph.

We can achieve this by creating a NumPy array with all zero values for the known size of the image and one channel for each bounding box.

Each bounding box is defined as min and max, x and y coordinates of the box.

These can be used directly to define row and column ranges in the array that can then be marked as 1.

All objects have the same class in this dataset. We can retrieve the class index via the ‘class_names‘ dictionary, then add it to a list to be returned alongside the masks.

Tying this together, the complete load_mask() function is listed below.

Finally, we must implement the image_reference() function.

This function is responsible for returning the path or URL for a given ‘image_id‘, which we know is just the ‘path‘ property on the ‘image info‘ dict.

And that’s it. We have successfully defined a Dataset object for the mask-rcnn library for our Kangaroo dataset.

The complete listing of the class and creating a train and test dataset is provided below.

Running the example successfully loads and prepares the train and test dataset and prints the number of images in each.

Now that we have defined the dataset, let’s confirm that the images, masks, and bounding boxes are handled correctly.

Test KangarooDataset Object

The first useful test is to confirm that the images and masks can be loaded correctly.

We can test this by creating a dataset and loading an image via a call to the load_image() function with an image_id, then load the mask for the image via a call to the load_mask() function with the same image_id.

Next, we can plot the photograph using the Matplotlib API, then plot the first mask over the top with an alpha value so that the photograph underneath can still be seen

The complete example is listed below.

Running the example first prints the shape of the photograph and mask NumPy arrays.

We can confirm that both arrays have the same width and height and only differ in terms of the number of channels. We can also see that the first photograph (e.g. image_id=0) in this case only has one mask.

A plot of the photograph is also created with the first mask overlaid.

In this case, we can see that one kangaroo is present in the photo and that the mask correctly bounds the kangaroo.

Photograph of Kangaroo With Object Detection Mask Overlaid

We could repeat this for the first nine photos in the dataset, plotting each photo in one figure as a subplot and plotting all masks for each photo.

Running the example shows that photos are loaded correctly and that those photos with multiple objects correctly have separate masks defined.

Plot of First Nine Photos of Kangaroos in the Training Dataset With Object Detection Masks

Another useful debugging step might be to load all of the ‘image info‘ objects in the dataset and print them to the console.

This can help to confirm that all of the calls to the add_image() function in the load_dataset() function worked as expected.

Running this code on the loaded training dataset will then show all of the ‘image info‘ dictionaries, showing the paths and ids for each image in the dataset.

Finally, the mask-rcnn library provides utilities for displaying images and masks. We can use some of these built-in functions to confirm that the Dataset is operating correctly.

For example, the mask-rcnn library provides the mrcnn.visualize.display_instances() function that will show a photograph with bounding boxes, masks, and class labels. This requires that the bounding boxes are extracted from the masks via the extract_bboxes() function.

For completeness, the full code listing is provided below.

Running the example creates a plot showing the photograph with the mask for each object in a separate color.

The bounding boxes match the masks exactly, by design, and are shown with dotted outlines. Finally, each object is marked with the class label, which in this case is ‘kangaroo‘.

Photograph Showing Object Detection Masks, Bounding Boxes, and Class Labels

Now that we are confident that our dataset is being loaded correctly, we can use it to fit a Mask R-CNN model.

How to Train Mask R-CNN Model for Kangaroo Detection

A Mask R-CNN model can be fit from scratch, although like other computer vision applications, time can be saved and performance can be improved by using transfer learning.

The Mask R-CNN model pre-fit on the MS COCO object detection dataset can be used as a starting point and then tailored to the specific dataset, in this case, the kangaroo dataset.

The first step is to download the model file (architecture and weights) for the pre-fit Mask R-CNN model. The weights are available from the GitHub project and the file is about 250 megabytes.

Download the model weights to a file with the name ‘mask_rcnn_coco.h5‘ in your current working directory.

• Download Weights (mask_rcnn_coco.h5) 246M

Next, a configuration object for the model must be defined.

This is a new class that extends the mrcnn.config.Config class and defines properties of both the prediction problem (such as name and the number of classes) and the algorithm for training the model (such as the learning rate).

The configuration must define the name of the configuration via the ‘NAME‘ attribute, e.g. ‘kangaroo_cfg‘, that will be used to save details and models to file during the run. The configuration must also define the number of classes in the prediction problem via the ‘NUM_CLASSES‘ attribute. In this case, we only have one object type of kangaroo, although there is always an additional class for the background.

Finally, we must define the number of samples (photos) used in each training epoch. This will be the number of photos in the training dataset, in this case, 131.

Tying this together, our custom KangarooConfig class is defined below.

Next, we can define our model.

This is achieved by creating an instance of the mrcnn.model.MaskRCNN class and specifying the model will be used for training via setting the ‘mode‘ argument to ‘training‘.

The ‘config‘ argument must also be specified with an instance of our KangarooConfig class.

Finally, a directory is needed where configuration files can be saved and where checkpoint models can be saved at the end of each epoch. We will use the current working directory.

Next, the pre-defined model architecture and weights can be loaded. This can be achieved by calling the load_weights() function on the model and specifying the path to the downloaded ‘mask_rcnn_coco.h5‘ file.

The model will be used as-is, although the class-specific output layers will be removed so that new output layers can be defined and trained. This can be done by specifying the ‘exclude‘ argument and listing all of the output layers to exclude or remove from the model after it is loaded. This includes the output layers for the classification label, bounding boxes, and masks.

Next, the model can be fit on the training dataset by calling the train() function and passing in both the training dataset and the validation dataset. We can also specify the learning rate as the default learning rate in the configuration (0.001).

We can also specify what layers to train. In this case, we will only train the heads, that is the output layers of the model.

We could follow this training with further epochs that fine-tune all of the weights in the model. This could be achieved by using a smaller learning rate and changing the ‘layer’ argument from ‘heads’ to ‘all’.

The complete example of training a Mask R-CNN on the kangaroo dataset is listed below.

This may take some time to execute on the CPU, even with modern hardware. I recommend running the code with a GPU, such as on Amazon EC2, where it will finish in about five minutes on a P3 type hardware.

Running the example will report progress using the standard Keras progress bars.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

We can see that there are many different train and test loss scores reported for each of the output heads of the network. It can be quite confusing as to which loss to pay attention to.

In this example where we are interested in object detection instead of object segmentation, I recommend paying attention to the loss for the classification output on the train and validation datasets (e.g. mrcnn_class_loss and val_mrcnn_class_loss), as well as the loss for the bounding box output for the train and validation datasets (mrcnn_bbox_loss and val_mrcnn_bbox_loss).

A model file is created and saved at the end of each epoch in a subdirectory that starts with ‘kangaroo_cfg‘ followed by random characters.

A model must be selected for use; in this case, the loss continues to decrease for the bounding boxes on each epoch, so we will use the final model at the end of the run (‘mask_rcnn_kangaroo_cfg_0005.h5‘).

Copy the model file from the config directory into your current working directory. We will use it in the following sections to evaluate the model and make predictions.

The results suggest that perhaps more training epochs could be useful, perhaps fine-tuning all of the layers in the model; this might make an interesting extension to the tutorial.

Next, let’s look at evaluating the performance of this model.

How to Evaluate a Mask R-CNN Model

The performance of a model for an object recognition task is often evaluated using the mean absolute precision, or mAP.

We are predicting bounding boxes so we can determine whether a bounding box prediction is good or not based on how well the predicted and actual bounding boxes overlap. This can be calculated by dividing the area of the overlap by the total area of both bounding boxes, or the intersection divided by the union, referred to as “intersection over union,” or IoU. A perfect bounding box prediction will have an IoU of 1.

It is standard to assume a positive prediction of a bounding box if the IoU is greater than 0.5, e.g. they overlap by 50% or more.

Precision refers to the percentage of the correctly predicted bounding boxes (IoU > 0.5) out of all bounding boxes predicted. Recall is the percentage of the correctly predicted bounding boxes (IoU > 0.5) out of all objects in the photo.

As we make more predictions, the recall percentage will increase, but precision will drop or become erratic as we start making false positive predictions. The recall (x) can be plotted against the precision (y) for each number of predictions to create a curve or line. We can maximize the value of each point on this line and calculate the average value of the precision or AP for each value of recall.

Note: there are variations on how AP is calculated, e.g. the way it is calculated for the widely used PASCAL VOC dataset and the MS COCO dataset differ.

The average or mean of the average precision (AP) across all of the images in a dataset is called the mean average precision, or mAP.

The mask-rcnn library provides a mrcnn.utils.compute_ap to calculate the AP and other metrics for a given images. These AP scores can be collected across a dataset and the mean calculated to give an idea at how good the model is at detecting objects in a dataset.

First, we must define a new Config object to use for making predictions, instead of training. We can extend our previously defined KangarooConfig to reuse the parameters. Instead, we will define a new object with the same values to keep the code compact. The config must change some of the defaults around using the GPU for inference that are different from how they are set for training a model (regardless of whether you are running on the GPU or CPU).