How to Train an Object Detection Model with Keras

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)

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.

Want Results with Deep Learning for Computer Vision?

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

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

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

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

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.

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).

Next, we can define the model with the config and set the ‘mode‘ argument to ‘inference‘ instead of ‘training‘.

Next, we can load the weights from our saved model.

We can do that by specifying the path to the model file. In this case, the model file is ‘mask_rcnn_kangaroo_cfg_0005.h5‘ in the current working directory.

Next, we can evaluate the model. This involves enumerating the images in a dataset, making a prediction, and calculating the AP for the prediction before predicting a mean AP across all images.

First, the image and ground truth mask can be loaded from the dataset for a given image_id. This can be achieved using the load_image_gt() convenience function.

Next, the pixel values of the loaded image must be scaled in the same way as was performed on the training data, e.g. centered. This can be achieved using the mold_image() convenience function.

The dimensions of the image then need to be expanded one sample in a dataset and used as input to make a prediction with the model.

Next, the prediction can be compared to the ground truth and metrics calculated using the compute_ap() function.

The AP values can be added to a list, then the mean value calculated.

Tying this together, the evaluate_model() function below implements this and calculates the mAP given a dataset, model and configuration.

We can now calculate the mAP for the model on the train and test datasets.

The full code listing is provided below for completeness.

Running the example will make a prediction for each image in the train and test datasets and calculate the mAP for each.

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.

A mAP above 90% or 95% is a good score. We can see that the mAP score is good on both datasets, and perhaps slightly better on the test dataset, instead of the train dataset.

This may be because the dataset is very small, and/or because the model could benefit from further training.

Now that we have some confidence that the model is sensible, we can use it to make some predictions.

How to Detect Kangaroos in New Photos

We can use the trained model to detect kangaroos in new photographs, specifically, in photos that we expect to have kangaroos.

First, we need a new photo of a kangaroo.

We could go to Flickr and find a random photo of a kangaroo. Alternately, we can use any of the photos in the test dataset that were not used to train the model.

We have already seen in the previous section how to make a prediction with an image. Specifically, scaling the pixel values and calling model.detect(). For example:

Let’s take it one step further and make predictions for a number of images in a dataset, then plot the photo with bounding boxes side-by-side with the photo and the predicted bounding boxes. This will provide a visual guide to how good the model is at making predictions.

The first step is to load the image and mask from the dataset.

Next, we can make a prediction for the image.

Next, we can create a subplot for the ground truth and plot the image with the known bounding boxes.

We can then create a second subplot beside the first and plot the first, plot the photo again, and this time draw the predicted bounding boxes in red.

We can tie all of this together into a function that takes a dataset, model, and config and creates a plot of the first five photos in the dataset with ground truth and predicted bound boxes.

The complete example of loading the trained model and making a prediction for the first few images in the train and test datasets is listed below.

Running the example first creates a figure showing five photos from the training dataset with the ground truth bounding boxes, with the same photo and the predicted bounding boxes alongside.

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 the model has done well on these examples, finding all of the kangaroos, even in the case where there are two or three in one photo. The second photo down (in the right column) does show a slip-up where the model has predicted a bounding box around the same kangaroo twice.

Plot of Photos of Kangaroos From the Training Dataset With Ground Truth and Predicted Bounding Boxes

Plot of Photos of Kangaroos From the Training Dataset With Ground Truth and Predicted Bounding Boxes

A second figure is created showing five photos from the test dataset with ground truth bounding boxes and predicted bounding boxes.

These are images not seen during training, and again, in each photo, the model has detected the kangaroo. We can see that in the case of the second last photo that a minor mistake was made. Specifically, the same kangaroo was detected multiple times.

No doubt these differences can be ironed out with more training, perhaps with a larger dataset and/or data augmentation, to encourage the model to detect people as background and to detect a given kangaroo once only.

Plot of Photos of Kangaroos From the Training Dataset With Ground Truth and Predicted Bounding Boxes

Plot of Photos of Kangaroos From the Training Dataset With Ground Truth and Predicted Bounding Boxes

Further Reading

This section provides more resources on the topic if you are looking to go deeper.






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

Specifically, you learned:

  • 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.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.

Develop Deep Learning Models for Vision Today!

Deep Learning for Computer Vision

Develop Your Own Vision Models in Minutes

...with just a few lines of python code

Discover how in my new Ebook:
Deep Learning for Computer Vision

It provides self-study tutorials on topics like:
classification, object detection (yolo and rcnn), face recognition (vggface and facenet), data preparation and much more...

Finally Bring Deep Learning to your Vision Projects

Skip the Academics. Just Results.

See What's Inside

662 Responses to How to Train an Object Detection Model with Keras

  1. Avatar
    Milemi June 1, 2019 at 5:38 am #

    Great tutorial !
    Could you give us advice how to annotate images, please ?
    What is the best practice ?
    How many images per object is enough ?
    How to annotate when there are several objects in the same image and they overlap ?
    Thank you.

    • Avatar
      Jason Brownlee June 1, 2019 at 6:17 am #

      Great questions, thanks!

      I hope to cover the topic in the future.

      • Avatar
        Yajuan October 15, 2020 at 7:40 pm #

        Dear Dr. Jason,

        I am a student from China. I am dealing with a problem related to scene classification and wondering if you could provide some good methods and materials.

        best, looking forward to hearing from you, Thank you for your time.

        Warm regards,
        Yajuan Xu

        • Avatar
          Jason Brownlee October 16, 2020 at 5:52 am #

          Thanks for the suggestion, I hope to write about the topic in the future.

    • Avatar
      Usama Ahmed October 7, 2019 at 8:40 pm #

      Here is the image annotation tool.

    • Avatar
      waleed June 4, 2020 at 11:51 pm #

      HEY JASON I need help in my satellite building images dataset I have labels in JSON format in which image coordinates in polygon shapes more than 4 points so mask rcnn is suitable for this kind of dataset because RPN needs 4 points to make a box but I have more than 4 points in my labels and annotated images as well so how it works for polygon? please help is there any method to convert polygon to 4 coordinates or any function which can help.

    • Avatar
      harsh verma October 9, 2021 at 8:20 pm #

      use labelimg repo

  2. Avatar
    simonYU June 4, 2019 at 6:05 pm #

    hi, Jason, while display_instances:

    running : display_instances(image, bbox, mask, class_ids, train_set.class_names)

    An error ocurred while starting the kernel ,

    home/user/anaconda3/bin/python: symbol lookup error: /home/user/anaconda3/lib/python3.6/site‑packages/numpy/core/../../../../ undefined symbol: __kmpc_global_thread_num

    Pls find the solution .

  3. Avatar
    roopesh June 4, 2019 at 6:19 pm #

    very nice steps !! How to predict with real time video (CCTV) instead of images, Thanks.

    • Avatar
      Jason Brownlee June 5, 2019 at 8:33 am #

      Great suggestion, I hope to cover it in the future.

    • Avatar
      Usama Ahmed September 26, 2019 at 11:59 pm #

      Use OpenCV to capture video from attached camera.

  4. Avatar
    maryam June 8, 2019 at 5:02 am #

    Hi Jason,
    Thank you very much for the precious tutorial. I face a problem in people counting project when I am going to track people though detecting them is not hard.
    would you please give me a tutorial about the best tracking methods such as “deep tracking” or other else?

  5. Avatar
    gary June 19, 2019 at 8:51 pm #

    Thank you very much for such a beautiful yet detailed tutorial. Its been great learning from you.

    • Avatar
      Jason Brownlee June 20, 2019 at 8:31 am #

      Thanks, I’m glad it helped.

      • Avatar
        gary June 24, 2019 at 10:27 pm #

        Hi jason, i am trying to train multiple object, how can i change the code to import multiple classes?
        Do i use multiple lines of:
        self.add_class(“dataset”, 1, “kangaroo”)
        self.add_class(“dataset”, 2, “tiger”)?

        • Avatar
          Jason Brownlee June 25, 2019 at 6:20 am #

          You can specify all of your classes with a unique integer.

          • Avatar
            Romell Domínguez August 28, 2019 at 6:47 am #

            Hi Jason, and then just add each image using:
            self.add_class(“dataset”, 1, “kangaroo”)
            self.add_class(“dataset”, 2, “tiger”)?

            self.add_image(‘dataset’, … )
            what parameter do i need set for identify ‘the class’

          • Avatar
            Romell Domínguez August 30, 2019 at 8:25 pm #

            i solve that problem for polygons shapes

          • Avatar
            Jason Brownlee August 31, 2019 at 6:04 am #

            I’m happy to hear that, well done!

          • Avatar
            Biki November 8, 2019 at 12:28 am #

            Please let me know how to do this

          • Avatar
            Akshay February 12, 2020 at 3:01 am #

            If we have both kangaroo and tiger inside single image, then how van I load the mask?

            self.add_class(“dataset”, 1, “kangaroo”)
            self.add_class(“dataset”, 2, “tiger”)

            I meant this part!!!

          • Avatar
            Jason Brownlee February 12, 2020 at 5:50 am #

            Sorry, I cannot review/debug your code.

  6. Avatar
    marry June 20, 2019 at 12:27 pm #

    ValueError: Dimension 1 in both shapes must be equal, but are 8 and 16. Shapes are [1024,8] and [1024,16]. for ‘Assign_682’ (op: ‘Assign’) with input shapes: [1024,8], [1024,16].

    hello,Jason,How to solve this error when calculating the mAP value?

  7. Avatar
    mahmoud July 10, 2019 at 6:30 pm #

    hi jason,

    I want to inquire about this file ~~mask_rcnn_kangaroo_cfg_0005.h5 ,
    how i can find it also why you seprate the training and predicting
    ,I mean at the last version of file it contains only the predicting with out the training ,how the model have saved the new weights after training so it can be used on the predicting step

    • Avatar
      Jason Brownlee July 11, 2019 at 9:46 am #

      The model is fit on the training dataset, saved, loaded and used to make prediction on a hold out test dataset.

      Does that help?

      • Avatar
        mahmoud July 11, 2019 at 6:18 pm #

        ya but my question befor train on dataset kangaroo i load weights to model
        # load weights (mscoco) and exclude the output layers
        model.load_weights(‘mask_rcnn_coco.h5’, by_name=True, exclude=[“mrcnn_class_logits”, “mrcnn_bbox_fc”, “mrcnn_bbox”, “mrcnn_mask”])
        then after training
        # load model weights
        model.load_weights(‘mask_rcnn_kangaroo_cfg_0005.h5’, by_name=True)
        why we load the weights again

        i have the file mask_rcnn_coco.h5, i think it have any initial weights ,but i do not know what is the file mask_rcnn_kangaroo_cfg_0005.h5 contains and where i can find this problem

        • Avatar
          Jason Brownlee July 12, 2019 at 8:32 am #

          The new set of weights is focused on only detecting kangaroos based on our own dataset.

          Does that help?

          • Avatar
            mahmoud July 12, 2019 at 5:58 pm #

            ya but i can not find this new set of weights ,i mean when it creats the file mask_rcnn_kangaroo_cfg_0005.h5

          • Avatar
            Jason Brownlee July 13, 2019 at 6:52 am #

            It will be in the same directory as the python file.

          • Avatar
            mahmoud July 14, 2019 at 7:25 am #

            thnx for your response ,another question how i can prepare my images to be on same structure of Kangaroo dataset to train and apply the model on it

          • Avatar
            Jason Brownlee July 14, 2019 at 8:18 am #

            It is not required, but it might be a helpful start if you are having trouble.

          • Avatar
            mahmoud eltaher July 15, 2019 at 7:27 pm #

            ya I need to do this because I want to implement the model on my problem so I have some images with some circles and I want to detect these circles

  8. Avatar
    Wolverin July 13, 2019 at 3:29 am #

    same problem with me, i am using google colab.

    this ‘mask_rcnn_kangaroo_cfg_0005.h5’ file is created while training as said in the blog. but i cannot find anywhere in my gdrive.

    • Avatar
      Jason Brownlee July 13, 2019 at 7:00 am #

      Perhaps try running on your workstation from the command line?

    • Avatar
      kevi October 8, 2019 at 1:03 am #

      I have ran on Google colab and .h5 saved in (/content/Mask_RCNN/kangaroo_cfg*/) folder. Check

    • Avatar
      zakaria October 10, 2019 at 3:40 am #

      Hi plz i am using colab google, i am having trouble installing RCNN Librery in python

      • Avatar
        Jason Brownlee October 10, 2019 at 7:03 am #

        I don’t know about colab, sorry.

        Perhaps try posting on stackoverflow?

  9. Avatar
    Jeremy Immanuel Putra Tandjung July 16, 2019 at 2:25 pm #

    Hello Jason,

    First of all, nice tutorial! Having the overall code at the end of each step really helped keep track of where I am in the code! Keep up the good job!

    I have a question, I notice that it took you on average a minute per epoch to train. However, I tried doing this with a different dataset and right now i’m on my first epoch and it’s ETA 3.5 hours. My desktop is fairly fast with a ryzen 7 cpu and a nvidia 1050Ti gpu.

    So is there something that I’m missing? My training dataset consist of 296 pictures of playing cards in different situations with a total file size of 30.4 MB (I’m trying to train a model to detect playing cards)

    Or is that a normal? Or is there some setting I’m missing?


    • Avatar
      Jason Brownlee July 17, 2019 at 8:15 am #

      It may be a factor of the number of images?

      It may be hardware?

      Perhaps experiment on some p3 EC2 instances or with a smaller dataset?

  10. Avatar
    Choi July 16, 2019 at 4:51 pm #

    Hi Jason.
    This post is so helpful to me to learn R-CNN training!

    As I do my work, I encounter some problems now.
    First I train the model based on ‘mask_rcnn_coco.h5’ weight first.

    So i got the model weight : ‘mask_rcnn_carpk_cfg_0010.h5’ file
    how can i append more training images and train based on above file?

    I just tried to append more images by load_images function, and next I trained the model by load_weights(‘mask_rcnn_carpk_cfg_0010.h5’, by_name=True, exclude=[“mrcnn_class_logits”, “mrcnn_bbox_fc”, “mrcnn_bbox”, “mrcnn_mask”])
    But it did not work..

    Is there any other things to set??

    Thank you!!

  11. Avatar
    Nathan Starliper July 17, 2019 at 5:33 am #

    Hi Jason,

    Great tutorial. However, I am bit confused as to why you used Mask RCNN instead of Faster RCNN? Mask RCNN is essentially Faster RCNN except with segmentation added. Here in this example you basically converted the segmentation into bounding boxes so it seems to me that it would have saved you quite a bit of effort and manual labor to just use Faster RCNN model instead?


    • Avatar
      Jason Brownlee July 17, 2019 at 8:31 am #

      Good question.

      Optionality. We can do object detection which is what most people want, with ability to do segmentation if needed.

  12. Avatar
    SATYAM SAREEN July 22, 2019 at 7:43 pm #

    Great Tutorial Sir,
    I really learned a lot.
    I have a doubt regarding multiclass detection. I have 2 classes: person with a helmet, person without a helmet. what changes should I make in the program? Like adding classes through add_class function.
    Huge Respect and Love.
    Satyam Sareen

    • Avatar
      Jason Brownlee July 23, 2019 at 8:00 am #

      Perhaps this tutorial will help you train your model:

      • Avatar
        SATYAM SAREEN July 23, 2019 at 7:38 pm #

        Good Afternoon Sir,

        You have attached the link to the same blog. Can you suggest the changes to be made in your code so that it runs smoothly for multiclass object detection?

        Warm Regards
        Satyam Sareen

        • Avatar
          Jason Brownlee July 24, 2019 at 7:52 am #

          What do you mean smoothly?

        • Avatar
          AutoRoboCulture November 18, 2019 at 3:05 am #

          Hello Satyam Sareen,

          Check out the code below, I have changed it to your requirement. If any query comment it down. Keep it up!


          class KangarooDataset(Dataset):
          # load the dataset definitions
          def load_dataset(self, dataset_dir, is_train=True):
          # define two class
          self.add_class(“dataset”, 1, “personWithHelmet”) #Change required
          self.add_class(“dataset”, 2, “personWithoutHelmet”) #Change required
          # define data locations
          images_dir = dataset_dir + ‘/images/’
          annotations_dir = dataset_dir + ‘/annots/’
          # find all images

          for filename in listdir(images_dir):
          # extract image id
          image_id = filename[:-4]
          #print(‘IMAGE ID: ‘,image_id)
          # skip all images after 90 if we are building the train set
          if is_train and int(image_id) >= 90: #set limit for your train and test set
          # skip all images before 90 if we are building the test/val set
          if not is_train and int(image_id) < 90:
          img_path = images_dir + filename
          ann_path = annotations_dir + image_id + '.xml'
          # add to dataset
          self.add_image('dataset', image_id=image_id, path=img_path, annotation=ann_path, class_ids = [0,1,2]) # for your case it is 0:BG, 1:PerWithHel.., 2:PersonWithoutHel… #Change required

          # extract bounding boxes from an annotation file
          def extract_boxes(self, filename):
          # load and parse the file
          tree = ElementTree.parse(filename)
          # get the root of the document
          root = tree.getroot()
          # extract each bounding box
          boxes = list()
          #for box in root.findall('.//bndbox'):
          for box in root.findall('.//object'): #Change required
          name = box.find('name').text #Change required
          xmin = int(box.find('./bndbox/xmin').text)
          ymin = int(box.find('./bndbox/ymin').text)
          xmax = int(box.find('./bndbox/xmax').text)
          ymax = int(box.find('./bndbox/ymax').text)
          #coors = [xmin, ymin, xmax, ymax, name]
          coors = [xmin, ymin, xmax, ymax, name] #Change required
          # extract image dimensions
          width = int(root.find('.//size/width').text)
          height = int(root.find('.//size/height').text)
          return boxes, width, height

          # load the masks for an image
          def load_mask(self, image_id):
          # get details of image
          info = self.image_info[image_id]
          # define box file location
          path = info['annotation']
          # load XML
          boxes, w, h = self.extract_boxes(path)
          # create one array for all masks, each on a different channel
          masks = zeros([h, w, len(boxes)], dtype='uint8')
          # create masks
          class_ids = list()
          for i in range(len(boxes)):
          box = boxes[i]
          row_s, row_e = box[1], box[3]
          col_s, col_e = box[0], box[2]
          if (box[4] == 'personWithHelmet'): #Change required #change this to your .XML file
          masks[row_s:row_e, col_s:col_e, i] = 2 #Change required #assign number to your class_id
          class_ids.append(self.class_names.index('personWithHelmet')) #Change required
          masks[row_s:row_e, col_s:col_e, i] = 1 #Change required
          class_ids.append(self.class_names.index('personWithoutHelmet')) #Change required

          return masks, asarray(class_ids, dtype='int32')

          # load an image reference
          def image_reference(self, image_id):
          info = self.image_info[image_id]
          return info['path']

          # define a configuration for the model
          class KangarooConfig(Config):
          # define the name of the configuration
          NAME = "kangaroo_cfg"
          # number of classes (background + personWithoutHelmet + personWithHelmet)
          NUM_CLASSES = 1 + 2 #Change required
          # number of training steps per epoch
          STEPS_PER_EPOCH = 90

          • Avatar
            Jason Brownlee November 18, 2019 at 6:50 am #

            Thanks for sharing.

          • Avatar
            Ashutosh Srivastava February 11, 2020 at 10:21 pm #

            You are great @AutoRoboCulture.

          • Avatar
            Akshay February 12, 2020 at 3:22 am #

            Halo, I tried to train with multiple classes in a single image, I am gettting an error like this

            File “C:……….\lib\site-packages\keras\engine\”, line 145, in standardize_input_data

            ValueError: Error when checking input: expected input_image to have shape (None, None, 1) but got array with shape (1024, 1024, 3).

            PS: I am working with gray scale images. and 3 classes. Inside single image both classes are present

          • Avatar
            Ademola Okerinde February 26, 2020 at 10:47 am #


          • Avatar
            Nourhan March 4, 2020 at 9:25 pm #

            Thank you so much for sharing these changes.
            However, after I followed all of them and adjusted the whole thing to fit my dataset, I keep getting this error:
            RuntimeError: generator raised StopIteration

            from that training line:
            model.train(train_set, test_set, learning_rate=config.LEARNING_RATE, epochs=5, layers=’heads’)

            Do you have any suggestions to overcome it?

          • Avatar
            hila March 12, 2020 at 10:14 pm #

            hey i was using your code for the training.
            can you please show us your prediction code? i was trying to use Jason’s but it came with a lot of errors which i cannot solve.

          • Avatar
            Andy August 2, 2020 at 1:36 am #

            Can you share the prediction part, please?

  13. Avatar
    mahmoud July 25, 2019 at 7:25 pm #

    is this model also suppose to detect the mask of the objects ,for the kangaroo on the images or we will need some modification to segment the images.

    • Avatar
      Jason Brownlee July 26, 2019 at 8:19 am #

      Yes, if masks are provided.

      In the case of kangaroos, we do not provide masks – just bounding boxes, therefore masks cannot be learned.

      • Avatar
        mahmoud July 29, 2019 at 8:47 pm #

        when i try to test image with multiple kangaroos ,it failed to detect them is there are two kangaroos interference it detect them as only one ?? any advice

        • Avatar
          Jason Brownlee July 30, 2019 at 6:11 am #

          Perhaps the model requires more training on photos with multiple kangaroos?

          • Avatar
            mahmoud July 30, 2019 at 10:02 pm #

            thanks for your response, another question is there a new version of Mask RCNN avilable on github .
            also what i need to have mask on my model how i can provide the model and make my model learn it also

          • Avatar
            Jason Brownlee July 31, 2019 at 6:52 am #

            The model can learn the mask, if you provide a dataset that has masks on the images.

          • Avatar
            mahmoud August 22, 2019 at 7:22 pm #

            thanks for your response i confused about some thing ,now we train model without mask ,so what is the mask loss on this case,and how it is calculated??

          • Avatar
            Jason Brownlee August 23, 2019 at 6:24 am #

            I don’t follow sorry, what do you mean exactly?

  14. Avatar
    Nishant Gaurav July 29, 2019 at 6:39 pm #

    I am getting this error. Please help
    OSError Traceback (most recent call last)
    in ()
    —-> 1 model.load_weights(‘mask_rcnn_kangaroo_cfg_0005.h5’, by_name=True)

    2 frames
    /usr/local/lib/python3.6/dist-packages/h5py/_hl/ in make_fid(name, mode, userblock_size, fapl, fcpl, swmr)
    140 if swmr and swmr_support:
    141 flags |= h5f.ACC_SWMR_READ
    –> 142 fid =, flags, fapl=fapl)
    143 elif mode == ‘r+’:
    144 fid =, h5f.ACC_RDWR, fapl=fapl)

    h5py/_objects.pyx in h5py._objects.with_phil.wrapper()

    h5py/_objects.pyx in h5py._objects.with_phil.wrapper()

    h5py/h5f.pyx in

    OSError: Unable to open file (unable to open file: name = ‘mask_rcnn_kangaroo_cfg_0005.h5’, errno = 2, error message = ‘No such file or directory’, flags = 0, o_flags = 0)

    • Avatar
      Jason Brownlee July 30, 2019 at 6:06 am #

      The error suggests that the path to your data file is incorrect or the file is corrupted in some way?

      • Avatar
        Nishant Gaurav July 31, 2019 at 9:45 pm #

        Thanks for the suggestion. The problem was resolved.
        How do we resolve the problem with the multiclass label? If we have to identify numbers and characters given in the same image and want to label all the characters and images, then how do we apply the multiclass label.

        • Avatar
          Jason Brownlee August 1, 2019 at 6:49 am #

          Perhaps extract the images of detected numbers (called segmentation), then classify each segmented image.

    • Avatar
      Mark November 10, 2020 at 1:59 pm #

      Hi I had the same “could not find file” problem. May you share how you resolve the problem?

  15. Avatar
    Dicko July 29, 2019 at 8:13 pm #

    Hi there, when I copied the example exactly, I am getting a train mAP of 0.000 and a test mAP of 0.000 also. Clearly something is wrong, I was wondering if anyone knew what the issue could be and how to resolve it. Thank you.

  16. Avatar
    saka July 30, 2019 at 2:26 pm #

    Dear Jason, Thanks! I really learned a lot.

    I am getting this error for the coding line “from mrcnn.utils import Dataset”.

    ” from mrcnn.utils import Dataset

    ModuleNotFoundError: No module named ‘mrcnn’ “.

    However, I checked if the library was installed by typing “show mask-rcnn” and got the results below,

    Name: mask-rcnn
    Version: 2.1
    Summary: Mask R-CNN for object detection and instance segmentation
    Author: Matterport
    License: MIT
    Location: c:\users\sakal\appdata\local\continuum\anaconda3\lib\site-packages\mask_rcnn-2.1-py3.7.egg

    According the information above, It seems no problem about the library installed. Could you please advise me about this. Thanks!!

  17. Avatar
    Dicko July 30, 2019 at 7:13 pm #

    Thanks for that I’ll have a look through the code and see if I’ve made a mistake somewhere when copying.
    Is there a file which has the complete code written so that i can just copy and past the whole lot rather than bits at a time?

    Thank you 🙂

    • Avatar
      Jason Brownlee July 31, 2019 at 6:48 am #

      Each of my tutorials has the complete file embedding, you can copy-paste it directly.

  18. Avatar
    Nishant Gaurav July 31, 2019 at 9:52 pm #

    File “”, line 21
    self.add_class(“dataset”, 2, “1”)
    IndentationError: unindent does not match any outer indentation level
    I am getting this error when I added just two new lines, in the code.

    def load_dataset(self, dataset_dir, is_train=True):
    # define one class
    self.add_class(“dataset”, 1, “N”)
    self.add_class(“dataset”, 2, “1”) //Added this new line
    # define data locations
    images_dir = dataset_dir + ‘/images/’
    annotations_dir = dataset_dir + ‘/annots/’

    for i in range(len(boxes)):
    box = boxes[i]
    row_s, row_e = box[1], box[3]
    col_s, col_e = box[0], box[2]
    masks[row_s:row_e, col_s:col_e, i] = 1
    class_ids.append(self.class_names.index(‘1′)) //Added this new line.
    return masks, asarray(class_ids, dtype=’int32’)

  19. Avatar
    Nishant Gaurav July 31, 2019 at 11:14 pm #

    IndexError Traceback (most recent call last)
    2 plt.imshow(image)
    3 # plot mask
    —-> 4 plt.imshow(mask[:, :, 0], cmap=’gray’, alpha=0.1)

    IndexError: index 0 is out of bounds for axis 2 with size 0

    I am getting this error after i added that extre two lines.

  20. Avatar
    ahmadreza August 1, 2019 at 2:44 am #

    hi Sir
    I am getting this error. Please help

    if is_train and int(image_id) >= 150:

    ValueError: invalid literal for int() with base 10: ‘Thumb’

  21. Avatar
    Nishant Gaurav August 1, 2019 at 2:16 pm #

    Hi Sir,
    Could you please give some insight where do I need to make changes for the multi-class label in the code so that I could identify the different characters and numbers in a single image?
    Please give some insight with examples so that it is easier to understand.
    Thanks so much for helping.

  22. Avatar
    mh August 6, 2019 at 6:58 pm #

    Thanks for your tutorial.

    But i want to ask is there any model can deal with the objects which have similar color on the back ground.

    • Avatar
      Jason Brownlee August 7, 2019 at 7:45 am #

      Perhaps. You may have to do some testing, or perhaps use transfer learning to tune an existing model.

  23. Avatar
    Tal August 9, 2019 at 12:33 am #

    Thank you very much for this great and clear tutorial!
    If I may ask:
    Is there a way to evaluate the model while training? For example at the end of each epoch?

    Thanks a million,


  24. Avatar
    Selman Bozkır August 14, 2019 at 4:57 am #

    Hi Jason,

    I have a problem. My dataset contains only 872 training images and 15 classes. Meanwhile, my images are rather bigger than kangroo or pascal voc files. They are around 1500 pixel wide and 1000 pixel tall. I have changed the python codes in order to apply multi-class classification. My equipment is 1050 ti on a 2