Tutorial 6 - Creating a more complex Annotator¶

Section 1 - Loading and using Machine Learning Models¶

In this tutorial you will create a more complex annotator that can detect objects and humans by using the a pretained Machine Learning Model of the YOLO family. YOLO stands for “You Only Look Once” and is a family of Machine Learning Models that only get to process an image once for object detection. This means that these models are relatively fast. The ultralytics library for python provides the functionality to load a pre-trained model from a file. We will create a new annotator called SimpleYoloAnnotator which you can find in src/robokudo/annotators/simple_yolo_annotator.py. As in the previous tutorials the update() and add_to_image() functions have to be implemented. All necessary libraries are already imported for you. The most important new type is the YOLO object. This object takes a single parameter representing the name of the model file. In this tutorial we will use the file "model.pt". The YOLO object will automatically load the model on initialization.

Task 6-1-1: Create a new class attribute self.model in the __init__() function. This should be a new YOLO object.
Task 6-1-2: In the update function load the COLOR_IMAGE from the CAS into a variable by creating a deepcopy of it.

The color image can be used to run “inference”, meaning the model gets the image data and returns its predictions to us. To do so efficiently we will use a context manager provided by another machine learning library called torch. torch is a lower level library that is also used by the ultralytics library internally. Add this code in the update function after loading the COLOR_IMAGE:

with torch.no_grad():

This context manager disables some features that are not needed in this tutorial and improves performance. The object in self.model is callable. This means to run inference with it we just have to call the model with a few parameters. For this tutorial pass your image as a positional argument and the keyword argument conf=0.9. The conf value is the minimum confidence percentage needed for a result to be accepted. Meaning a prediction will only be returned if the model is atleast 90% sure that the prediction is correct.

Task 6-1-3: Call the YOLO object inside of the context manager and write the results into a new variable for example results.

The new variable will now contain a list of Results objects. One object for every image we passed to the model. This means our list will only contain a single instance as only a single image was passed to the model.

The Results object itself has an attribute called boxes containing a of Boxes object. This object then contains the bounding boxes of the detected objects. However they are still stored as torch.Tensor objects which have to be converted to something we can access better. The Boxes class offers a few functions for this:

Boxes.cpu()

This will move the tensors to CPU memory for example in case they are stored on the GPU memory. We are not using a GPU, but we will call this anyway to be safe.

Boxes.numpy()

This will turn all tensors in the Boxes object into numpy arrays.

Task 6-1-4: Load the boxes object into cpu memory and convert its tensors to a numpy arrays. Store the result in a variable like results_np
Task 6-1-5: Create a loop over results_np an for every item do the following:
- Save the bounding box in a variable like bbox in the xyxy format by accessing item.xyxy[0]
- Add a white rectangle to the COLOR_IMAGE you loaded in the previous tasks using the datapoints in bbox (Note: You dont have to implement add_to_image yet)
- Write the new COLOR_IMAGE to the annotator outputs as done in the previous tutorials
Task 6-1-6: Create a new Sequence containing the SimpleYoloAnnotator. Make it run on the query type yolo-detect instead of the current sequence on that query type.

The code for the SimpleYoloAnnotator could look something like this:

# ...
class SimpleYoloAnnotator(robokudo.annotators.core.BaseAnnotator):

    def __init__(self, name="SimpleYoloAnnotator"):
        super().__init__(name=name)
        self.model = YOLO("best.pt")

    def update(self):
        # Read out the input color image and create a copy
        visualization_img = copy.deepcopy(self.get_cas().get(CASViews.COLOR_IMAGE))

        # Use the context manager to improve performance
        with torch.no_grad():
            # Run inference on the color image from the CAS
            result_tensor = self.model(visualization_img, conf=0.9, show=False)[0]

        # Move the Results object to cpu memory and convert its tensors to numpy arrays
        result_np = result_tensor.boxes.cpu().numpy()

        # Iterate over all results
        for result in result_np:
            # Read out the bounding box in xyxy format
            bbox = result.xyxy[0]
		   
            # Add the bounding box to the visualization_img
            visualization_img = cv2.rectangle(visualization_img, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (255, 255, 255), 2)

        # Write the final image containing the newly created rectangles to the
        # annotator output
        self.get_annotator_output_struct().set_image(visualization_img)
        return py_trees.Status.SUCCESS
# ...

Section 2 - Classifying Objects and creating Annotations¶

Now that the object detection is working, we will implement object classification. The best.pt model used exposes its available classes after loading in form of a dictionary with numbers as keys. These numbers are ids for the classes. These ids are returned by the inference, so this dictionary makes it possible to know what the actual class name is. In our case the dictionary can be accessed through the YOLO object we created in section one by using self.model.names.

Task 6-2-1: Create a new class attribute self.id2name in the __init__() function containing the id to class pairs.

For each of the results we got out of the inference we now want to create annotations. Since we are detecting objects the ObjectHypothesis will be used. This annotation however can not store classification data so we will also attach a robokudo.types.annotations.Classification to the ObjectHypothesis objects. The class of a result can be read out through result.cls[0], which returns the id that then has to be used with the newly created self.id2name to return a class in textual form. We can also read out how much confidence the model has in its prediction by reading out result.conf[0]. Every annotation in RoboKudo can also store its source annotator in the class field source. This is usually set to self.name from within the annotator and is useful in scenarios where you need to find out which annotator added the annotation.

Task 6-2-2: Inside of the results loop store the class id and class name into two separate variables.
Task 6-2-3: Create a new Classification object for each result and store the class name into its class field name and the confidence into its class field confidence. Also store the annotators source in the classes source field.

There is a lot of information that we can now put into the ObjectHypothesis. We already have the bounding box for each object and we also know its type. Because ObjectHypothesis inherit from the IdentifiableAnnotation we can also store and object id that can be used to uniquely identify an annotation. For now we will just use the index of the results list for this value. When there are multiple sources of ObjectHypothesis however, the id has to be retrieved in a different manner.

Task 6-2-4: Create a new ObjectHypothesis with the following data
- The class fields source and id as explained above
- The class field name should contain the textual class name
- Pass the bounding box to the roi.roi object in the ObjectHypothesis
- Add the Classification object to the ObjectHypothesis.annotations list

Tip

The results bounding box is currently in the xyxy form, but the roi.roi only has x, y, width and height.

Important

The bounding box coordinates are of type float, the roi.roi however only takes in the type int

Now that we have a complete ObjectHypothesis we have to add it to the CAS somehow so that other annotators can access these annotations too. To do so you can add the annotation to the self.get_cas().annotations list.

Task 6-2-5: Add the newly created ObjectHypothesis to the CAS

Tip

To make this a bit more efficient you can store the ObjectHypothesis in a small list during the loop and add them all at once to the CAS by using the extend function offered by lists.

Task 6-2-6: Send a new yolo-detect query to RoboKudo and confirm through the details tree in the CAS output visualization that all the values you added to the ObjectHypothesis are correct. This should include
- The objects source annotator
- The objects position, which is derived from roi.roi
- The objects id
- A Classification annotation with
  - The source annotator
  - The classname and class id
  - The confidence
Task 6-2-7: Implement the add_to_image function with the following critearia
- The bounding box should be added there instead of directly in the loop
- Add a text to the bounding box containing both the class name and confidence
Task 6-2-8: Send a new yolo-detect query to RoboKudo and confirm that the bounding boxes and text are correctly added to the output image.

The final output of the annotator could look something like this:

The final code for the SimpleYoloAnnotator could look something like this:

# ...
class SimpleYoloAnnotator(robokudo.annotators.core.BaseAnnotator):

    def __init__(self, name="SimpleYoloAnnotator"):
        super().__init__(name=name)
        self.model = YOLO("best.pt")
        self.id2name = self.model.names

    def update(self):
        visualization_img = copy.deepcopy(self.get_cas().get(CASViews.COLOR_IMAGE))

        with torch.no_grad():
            result_tensor = self.model(visualization_img, conf=0.9, show=False)[0]

        result_np = result_tensor.boxes.cpu().numpy()
        object_hypotheses = []
        for obj_id, result in enumerate(result_np):
            bbox = result.xyxy[0]

            cls = result.cls[0]
            name = self.id2name[cls]

            classification = robokudo.types.annotation.Classification()
            classification.source = self.name
            classification.classname = name
            classification.confidence = result.conf[0]

            object_hypothesis = robokudo.types.scene.ObjectHypothesis()
            object_hypothesis.source = self.name
            object_hypothesis.type = cls
            object_hypothesis.id = obj_id

            object_hypothesis.roi.roi.pos.x = int(bbox[0])
            object_hypothesis.roi.roi.pos.y = int(bbox[1])
            object_hypothesis.roi.roi.width = int(bbox[2] - bbox[0])
            object_hypothesis.roi.roi.height = int(bbox[3] - bbox[1])

            object_hypothesis.annotations.append(classification)
		   
            visualization_img = self.add_to_image(object_hypothesis, visualization_img)
            
            object_hypotheses.append(object_hypothesis)
        
        self.get_cas().annotations.extend(object_hypotheses)
        self.get_annotator_output_struct().set_image(visualization_img)
        return py_trees.Status.SUCCESS

    @staticmethod
    def add_to_image(obj, image):
        x1, y1, x2, y2 = (obj.roi.roi.pos.x, obj.roi.roi.pos.y,
                          obj.roi.roi.pos.x + obj.roi.roi.width,
                          obj.roi.roi.pos.y + obj.roi.roi.height)

        vis_text = f"{obj.annotations[0].classname}, {obj.annotations[0].confidence:.2f}"
        font = cv2.FONT_HERSHEY_COMPLEX
        image = cv2.putText(image, vis_text, (x1, y1 - 5), font, 0.5,
                            (0, 0, 255), 1, 2)
        image = cv2.rectangle(image, (x1, y1), (x2, y2), (255, 255, 255), 2)
        return image

Section 3 - Searching for objects¶

Detecting objects is already a good start. However it is even more useful to be able to search for objects aswell. As already done in the previous tutorials we will now implement another small annotator called ClassFilterAnnotator. As the name suggests it is again used to filter objects depending on the query that was sent to RoboKudo. This time we will use the obj.type field to search for the different classes however. You can find the template annotator in class_filter.py.

Implement the annotators update function so that the objects are added to the output image with the add_to_image function only if obj.type is equal to the objects classification.
Add the annotator to the sequence after the SimpleYoloAnnotator and try to identify the milk by sending a yolo-detect query while specifying a type.

The final output should look something like this:

The final code for the class_filter.py could look like this:

class ClassFilterAnnotator(robokudo.annotators.core.BaseAnnotator):
    # ...
    def update(self) -> py_trees.common.Status:
        query = self.get_cas().get(CASViews.QUERY)
        annotations = self.get_cas().filter_annotations_by_type(robokudo.types.scene.ObjectHypothesis)
        visualization_img = copy.deepcopy(self.get_cas().get(CASViews.COLOR_IMAGE))

        for annotation in annotations:
            for oh in annotation.annotations:
                if (isinstance(oh, robokudo.types.annotation.Classification)
                        and query.obj.type.lower() == oh.classname.lower()):
                    visualization_img = self.add_to_image(annotation, visualization_img)

        self.get_annotator_output_struct().set_image(visualization_img)
        return py_trees.Status.SUCCESS
    # ...