Output tensor of Yolo11

[spiderwei]

Could you please tell me where the object confidence score and class confidence score are in the output tensor of the yolo11 model?

[UltralyticsAssistant]

👋 Hello @spiderwei, thank you for reaching out to Ultralytics! 🚀

We recommend visiting our Docs for a comprehensive introduction, including examples of how to interpret YOLO model outputs and where you can find answers to commonly asked questions.

🌟 If your question relates to a bug or unexpected behavior, it would be very helpful if you could provide a minimum reproducible example to assist us in diagnosing the issue.

🤔 For questions about custom training or model output interpretation, please supply as much detail as you can. This includes snippets of your code, model configurations, and any outputs you are encountering.

Join the Ultralytics community where you feel most comfortable. For lively discussions and support, check us out on Discord 🎧. Prefer a forum format? Visit Discourse. For deeper dives and knowledge sharing, our Subreddit is another great avenue.

Upgrade

Ensure you have the latest version of ultralytics along with all requirements in a Python>=3.8 environment with PyTorch>=1.8:

pip install -U ultralytics

Environments

Run YOLO in these supported environments, with all necessary dependencies preinstalled:

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. Check the GCP Quickstart Guide
• Amazon Deep Learning AMI. See the AWS Quickstart Guide
• Docker Image. Check the Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing, verifying the correct operation of all YOLO Modes and Tasks across different platforms every 24 hours and with each commit.

This response is automated, but an Ultralytics engineer will assist you soon. 😊

[glenn-jocher]

@spiderwei the YOLO11 output tensor includes object confidence scores and class probabilities. The object confidence is typically found in the first channel, while class probabilities follow. For detailed tensor structure, refer to the model documentation.

[glenn-jocher]

@spiderwei in YOLO11, the object confidence score is typically located before the class probabilities in the output tensor, similar to YOLOv8. If you're encountering issues, please ensure you're using the correct model version and refer to the latest documentation for any updates.

[spiderwei]

Thank you for your reply. I have collected some data in the past few days, which are the output tensors of yolov8 and yolov5. I noticed that the output of yolov8 (batchsize, 84, 8400) is similar to the output of yolo11. But the "85" in the output of yolov5 (as shown below) contains the object confidence value, which is not available in the corresponding channels of the yolov8(yolo11) outputs. So I want to know where the target confidence missing from yolov8 and yolo11 is?

The first two pictures are the output tensor graphs of yolov8 and yolov5 that I collected, and the last one is the output tensor graph of yolo11 that I got through debugging (I set batchsize to 1).

[glenn-jocher]

@spiderwei in YOLO11, the object confidence is integrated into the class probabilities, similar to YOLOv8. The confidence score is derived from the combination of objectness and class probabilities, unlike YOLOv5 where it's explicitly separated.

[spiderwei]

What formula is used to combine object confidence and class probabilities into the class confidence score of yolo11? Can you tell me which file has the instructions? If you can help, I will be very grateful.

[glenn-jocher]

The class confidence score in YOLO11 is typically the product of the object confidence and class probabilities. For implementation details, please check the model's inference code in the models directory.

[vyrsh]

hello @glenn-jocher
I am sorry, I tried to find the documentation on https://docs.ultralytics.com/models/yolo11/ and https://docs.ultralytics.com.

but I was not able to find the model documentation.

I actually want to find the documentation for the output tensor. as I do not get what the 8400 stands for in the (1,84,8400).
I know that the output tensor for a yolo model usually considers x,y of the top left corner. the height and width of the box. the confidence score and the objections. but this is 6, not 8400.

thankyou for your assistance.

[glenn-jocher]

@vyrsh thank you for your question. The YOLO11 output tensor of shape (1, 84, 8400) can be understood as follows:

• 84: Represents the number of channels, where the first 4 are for bounding box parameters (x, y, width, height), the 5th is the object confidence score, and the remaining 80 correspond to class probabilities for COCO's 80 classes.
• 8400: Refers to the total number of grid points or anchor locations where predictions are made.

This structure enables YOLO to predict bounding boxes, object confidence, and class probabilities for multiple objects in an image. For further details, you can reference YOLO11 documentation. Let us know if you need additional clarification!