Introduction

In standard object detection, the output is typically a rectangular bounding box. However, in certain scenarios, objects may have a rotated shape, requiring a bounding box with a rotation angle. This type of bounding box is referred to as an OBB (Oriented Bounding Box).

Standard detection results: x, y, w, h represent the top-left or center coordinates of the rectangle, along with its width and height.
OBB detection results: x, y, w, h, angle includes an additional parameter for the rotation angle.

Using Rotated Bounding Boxes (OBB) in MaixPy MaixCAM

MaixPy supports the YOLO11/YOLOv8 OBB model, enabling quick and convenient implementation. Below is an example based on the nn_yolo11_obb.py script from MaixPy/examples:

from maix import camera, display, image, nn, app

detector = nn.YOLO11(model="/root/models/yolo11n_obb.mud", dual_buff=True)

cam = camera.Camera(detector.input_width(), detector.input_height(), detector.input_format())
disp = display.Display()

while not app.need_exit():
    img = cam.read()
    objs = detector.detect(img, conf_th=0.5, iou_th=0.45)
    for obj in objs:
        points = obj.get_obb_points()
        msg = f'{detector.labels[obj.class_id]}: {obj.score:.2f}, {obj.angle * 180:.1f}'
        img.draw_string(points[0], points[1] - 4, msg, color=image.COLOR_RED)
        detector.draw_pose(img, points, 8 if detector.input_width() > 480 else 4, image.COLOR_RED, close=True)
    disp.show(img)

In this example, the YOLO11 model is used to load an OBB model. After detecting an object, the rotation angle can be accessed via obj.angle. The x, y, w, h attributes of obj represent the unrotated rectangle. The get_obb_points method retrieves the four vertices of the rotated rectangle, and draw_pose is used to draw the rotated bounding box. The close parameter ensures the vertices are connected.

The default model is the official YOLO11 15-class model with the following labels:

plane, ship, storage tank, baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, large vehicle, small vehicle, helicopter, roundabout, soccer ball field, swimming pool

The model file can be found at /root/models/yolo11n_obb.mud.

More Input Resolutions

The default input image resolution is 320x224. For higher resolutions, download from the MaixHub Model Zoo or customize your own model as described below.

Customizing Your Own OBB Model for MaixPy MaixCAM

Using the Model on a Computer

For an introduction to the official YOLO11 OBB models, refer to YOLO11 OBB. This documentation explains how to use OBB models on a computer and export ONNX model files.

Exporting the Model for MaixCAM

Follow the YOLO11/YOLOv8 Custom Models guide to convert an ONNX model into a MUD model compatible with MaixCAM.
Note: Ensure the output names of the model's layers are correctly configured during conversion.

Training Your Own OBB Model

Refer to the YOLO11 Official Training Documentation to prepare your dataset and train your own OBB model.