Oriented Bounding Boxes (YOLOv8-OBB) Object Detection using Ultralytics YOLOv8 | Episode 21

Join us on the 21st video of our new series as we explore the vast capabilities of Ultralytics YOLOv8 Oriented Bounding Boxes (YOLOv8-OBB) models for developing projects and applications.
In this episode, we will extensively examine the YOLOv8-OBB module, providing a thorough overview. We'll explore its practical applications and complexities, delving into the code implementation to foster a comprehensive understanding.
➡️ Explore more: docs.ultralytics.com/tasks/obb/
➡️ Bilibili Video - shorturl.at/ajlo2
Key Moments
0:00 - Ultralytics YOLOv8-OBB Introduction
0:38 - YOLOv8-OBB Documentation
1:26 - YOLOv8-OBB Models
1:45 - YOLOv8-OBB Train
2:13 - YOLOv8-OBB Dataset Format
2:47 - YOLOv8-OBB Val & Predict
3:05 - YOLOv8-OBB Export
3:37 - YOLOv8-OBB Inference using Python
5:23 - YOLOv8-OBB Demo
5:46 - YOLOv8-OBB Use Cases
6:05 - Summary
Ultralytics ⚡ resources
- About Us - ultralytics.com/about
- Join Our Team - ultralytics.com/work
- Contact Us - ultralytics.com/contact
- Discord - ultralytics.com/discord
- Ultralytics License - ultralytics.com/license
YOLOv8 🚀 resources
- GitHub - github.com/ultralytics/ultral...
- Docs - docs.ultralytics.com/

[@Vahroc]

Lets assume we crop a big image into smaller regions (kinda like a sliding window) such that sometimes half a bounding box (or at least one corner of a box) sit outside the newly cropped image. When normalizing between 0 and 1, should we set the coordinate of the four points normaly, such that one of the corner may have coordinates bigger than 1 or smaller than 0 ? Or should we adapt these values to fall into the cropped image ? But then the box is not necessarily rectangular anymore. What is the best approach to train with yolo OBB under these conditions ?

[@Ultralytics]

When training with YOLO OBB on cropped images, it's best to normalize bounding box coordinates based on the original image size, even if they fall outside the [0, 1] range. This maintains the integrity of the bounding boxes. Ensure your data pipeline and loss function can handle out-of-bound coordinates. If required to clip, adjust coordinates to fit within the cropped image while preserving box validity.

[@afjamo]

Cooool!

[@Ultralytics]

Thank you!!!

[@m033372]

Nice!

[@Ultralytics]

Thanks!

[@frostscarlet]

doest it uses the same approach as yolov5-obb which use CSL for its OBB function?

[@Ultralytics]

Well, Ultralytics YOLOv8 uses a completely different approach, which is the piou(it's like iou to regular bounding boxes but for rotated boxes) way to regress obb. Thanks

[@sudhikrishnana9778]

Can the OBB annotated images can be used with a mix of traditional bounding boxes images dataset for the training? Will it affect the output performance?

[@Ultralytics]

Oriented bounding boxes are a separate concept that involves angle calculations for improved results. You cannot directly use OBB with traditional bounding boxes, but with modifications, you may be able to integrate them with traditional bounding boxes. Thanks, Ultralytics Team!

[@TravelwithRasel.]

hey, as a learner, i have a question of a point, should we mention the name of source into the last line, i meant inside the result row, please try to clarify of my concern

[@Ultralytics]

Yes, you'll need to use the pathway to the video file or live stream for source argument, which will be use for inference. Thanks Ultralytics Team!

[@user-xq8gv9ep9c]

How can get the coordinates of the results image predicted by custom model OBB? because if use the none OBB code, the result getting None

[@Ultralytics]

You can obtain the bounding box coordinates for the Oriented Bounding Box (OBB) task using the provided code below. """python from ultralytics import YOLO from ultralytics.utils.plotting import Annotator, colors import cv2 model = YOLO("yolov8n-obb.pt") names = model.names cap = cv2.VideoCapture("Path/to/video/file.mp4") assert cap.isOpened(), "Error reading video file" while cap.isOpened(): success, im0 = cap.read() if success: results = model.track(im0, persist=True, show=False) pred_boxes = results[0].obb annotator = Annotator(im0, line_width=3, example=names) for d in reversed(pred_boxes): c, id = int(d.cls), None if d.id is None else int(d.id.item()) label = names[c] + ":" + str(id) box = d.xyxyxyxy.reshape(-1, 4, 2).squeeze() print("Bounding Box Coordinates : ", box) annotator.box_label(box, label, color=colors(int(id), True), rotated=True) cv2.imshow("ultralytics", im0) if cv2.waitKey(1) & 0xFF == ord('q'): break continue break cap.release() cv2.destroyAllWindows() """ Thanks

[@dalinsixtus6752]

for the cli command !yolo task=obb ... how to implement the same using model='best.pt' ,model.predict() , what is the arg for task=obb ??

[@Ultralytics]

Oriented bounding boxes use the same parameters as those employed by the official YOLOv8 model. For more info: docs.ultralytics.com/modes/predict/#inference-arguments Thanks Ultralytics Team!

[@dalinsixtus6752]

@@Ultralytics is it possible to change the bounding box color during live camera detection , if certain condition is satisfied after object detection i.e i need to change the color from green to red if certain condition is proved false using ultralytics.utils.....

[@Ultralytics]

Yes, it is achievable. Within the annotator.box_label function, you have the option to define the color of bounding boxes. For additional details, you can refer to our documentation at: docs.ultralytics.com/reference/utils/plotting/?h=box_la#ultralytics.utils.plotting.Annotator.box_label Thanks

[@dalinsixtus6752]

@@Ultralytics thanks for the solution.How to combine two models of different class , i tried transfer learning still didn't got the classes from the second model , even there is no module for ensembling yolov8 , can you provide me solutions to combine two models of different classes into a single model

[@zubairkhalid3209]

Is there a possibility to quantify the number of pixels that particularly define that bounding box (so basically quantifying sizes from say microscopic images)? Also, is it possible to quantify the the intensity of certain colors in a bounding box (just like colorimeter)?

[@Ultralytics]

Yes, it's possible to quantify the number of pixels defining a bounding box in YOLOv8, aiding size quantification in microscopic images. Additionally, YOLOv8 allows for quantifying color intensity within a bounding box, akin to a colorimeter, offering versatile analysis capabilities. Note: By default, Ultralytics YOLOv8 does not include support for these features, but you can customize and add them based on your specific use case. Thanks Ultralytics Team!

[@user-di4eb8vc1i]

how the dataset be converted from the yolov8 annotated format to the yolov8 oriented bounding boxes format?

[@Ultralytics]

At the moment, there isn't a direct method to convert an object detection dataset to OBB format. Our team is actively developing OBB modules, and you may see this conversion on our documentation page soon: docs.ultralytics.com/ Thanks, Ultralytics Team!

[@user-cs6ni1oo6s]

😍Any visual heatmaps adapted for OBB object detection?

[@Ultralytics]

Heatmaps for OBB object detection are currently not supported. However, we do plan to include this feature in upcoming releases. Thank you, Ultralytics Team

[@dimitrispolitikos1246]

Very nice work! A short question, which software tool do you suggest to use for annotating our dataset with oriented bounding boxes?

[@Ultralytics]

Thanks. To annotate the dataset in OBB format, we recommend using the LabelImg_OBB Annotation tool, which can be found at this GitHub link: github.com/heshameraqi/labelImg_OBB Best regards, Ultralytics Team

[@dimitrispolitikos1246]

@@Ultralytics thank you very much for the suggestion! I tried it in the past but I think that it returns the bounding boxes as "class_id, x,y, r, theta" and not "class_index, x1, y1, x2, y2, x3, y3, x4, y4". I will double check it. Maybe a conversion function will be needed. Warmest regards!!

[@truonggiang-227nguyen5]

how can I get the coordinates of the bounding box? Thanks you!

[@Ultralytics]

To obtain bounding box coordinates, access the output of Ultralytics YOLOv8 model. Typically, it provides (x, y, width, height, conf, id) values. Extract them from the detection results, adjusting the code based on your model's output format. Thanks

[@dinihanafi5097]

@@Ultralytics can you provide the example code?

[@Ultralytics]

Sure, Below is the provided code snippet for obtaining the coordinates of Oriented Bounding Boxes using Ultralytics YOLOv8. ```python from ultralytics import YOLO from ultralytics.utils.plotting import Annotator, colors import cv2 # Initialize YOLOv8 model model = YOLO("yolov8n-obb.pt") names = model.names # Open video file cap = cv2.VideoCapture("Path/to/video/file.mp4") assert cap.isOpened(), "Error reading video file" while cap.isOpened(): success, im0 = cap.read() if success: # Make predictions on each frame results = model.predict(im0, persist=True, show=False) pred_boxes = results[0].obb # Initialize Annotator for visualization annotator = Annotator(im0, line_width=2, example=names) # Iterate over predicted bounding boxes and draw on image for d in reversed(pred_boxes): box = d.xyxyxyxy.reshape(-1, 4, 2).squeeze() print("Bounding Box Coordinates : ", box) annotator.box_label(box, names[int(d.cls)], color=colors(int(d.cls), True), rotated=True) # Display annotated image cv2.imshow("ultralytics", im0) # Check for key press to exit if cv2.waitKey(1) & 0xFF == ord('q'): break continue break # Release video capture and close windows cap.release() cv2.destroyAllWindows() ``` Thanks

[@user-vo2kg3yp1j]

Nice, can you give me your test video(ships.mp4)

[@Ultralytics]

Sure, the video is available at: shorturl.at/gjvCE Thanks Ultralytics Team!

[@user-vo2kg3yp1j]

@@Ultralytics thanks