Ultralytics
zh/tasks/pose/ #8788
giscus[bot]
bot
Replies: 22 comments 38 replies
-
|
I have reproduced the result of yolov8s-pose, and the experimental data is very different from yours, with map50 dropping 20% points. My code and data are exactly the same as what you described, and the operation steps are also strictly followed by yours. Therefore, I may have a problem with hyperparameter setting. |
Beta Was this translation helpful? Give feedback.
-
|
@AliasChenYi hey there! 👋 It looks like you're experiencing a significant drop in mAP50 with the YOLOv8s-pose model compared to our reported results. Thanks for sharing your hyperparameter settings; they seem mostly in line with standard configurations. One potential area to explore is the Another aspect to consider is the Lastly, the If after adjusting these you're still seeing discrepancies, it might be helpful to double-check the dataset integrity and ensure there's no data leakage or preprocessing differences. Feel free to share more details if the issue persists! 🚀 |
Beta Was this translation helpful? Give feedback.
-
|
I already look.
…---Original---
From: ***@***.***>
Date: Fri, Mar 8, 2024 18:41 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: [Ultralytics] zh/tasks/pose/ (Discussion #8788)
zh/tasks/pose/
了解如何使用Ultralytics YOLOv8 完成姿势估计任务。查找预训练模型，学习如何训练、验证、预测和输出自己的模型。
https://docs.ultralytics.com/zh/tasks/pose/
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
@AliasChenYi 感谢您的关注和提问！🙌 看起来您已经查看了我们的姿势估计任务页面。如果您有任何具体问题或需要进一步的帮助，比如如何训练您自己的姿势估计模型，或者如何使用我们的预训练模型进行预测，请随时告诉我！ 例如，如果您想使用预训练的YOLOv8n-pose模型进行预测，您可以这样做： from ultralytics import YOLO
# 加载预训练模型
model = YOLO('yolov8n-pose.pt')
# 对图像进行预测
results = model('path/to/your/image.jpg')
# 显示结果
results.show()希望这有助于您开始！如果有任何其他问题，我在这里帮助您。🚀 |
Beta Was this translation helpful? Give feedback.
-
|
Hello, may I ask if you solved this problem? I'm also having this issue when training the coco pose dataset, it's about the same hyperparameter as you used, but in the end the mAP is reduced by 20% . I don't know which parameter is problematic. |
Beta Was this translation helpful? Give feedback.
-
|
@2823387967 hello! 👋 It sounds like you're encountering a challenge with training on the COCO pose dataset. When you notice a significant drop in mAP, it's often related to the dataset's complexity or the model's capacity to learn from it. Here are a couple of suggestions:
Here's a quick example of adjusting the learning rate and using a different model variant: yolo pose train data=coco128-pose.yaml model=yolov8s-pose.pt epochs=100 imgsz=640 lr0=0.005Remember, every dataset is unique, so it might take some experimentation to find the optimal settings. If you continue to face issues, feel free to share more details or reach out on our GitHub discussions for more targeted advice. Good luck! 🚀 |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for your reply, I used pre-training weights to reduce the learning rate, and found that the mAP increased by 10% compared to the previous experiment, but after 30 rounds the mAP increased very slowly, from 30 rounds to 100 rounds only increased by 3%.Is there a hyperparameter in YOLOv8 that can reduce the learning rate faster during training. |
Beta Was this translation helpful? Give feedback.
-
|
@2823387967 hey there! 🌟 It's great to hear that adjusting the learning rate with pre-training weights improved your mAP by 10%. Regarding the slow increase in mAP after 30 epochs, this is quite common as the model starts to converge. For adjusting the learning rate dynamically during training, YOLOv8 supports the yolo detect train data=coco128.yaml model=yolov8n.pt epochs=100 lr-schedule=cosineThis adjusts the learning rate following a cosine curve, which can help in reducing the learning rate more effectively as training progresses. Keep experimenting with different schedules to find what works best for your specific case. Happy training! 🚀 |
Beta Was this translation helpful? Give feedback.
-
|
At present, it has not been solved, I do not know what is the reason, debugging a lot of parameters can not be done.
…---Original---
From: ***@***.***>
Date: Mon, Mar 11, 2024 15:07 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
Hello, may I ask if you solved this problem? I'm also having this issue when training the coco pose dataset, it's about the same hyperparameter as you used, but in the end the mAP is reduced by 20% . I don't know which parameter is problematic.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
Thank you for your reply. I will continue to adjust the parameters and hope to make some improvements. After you solve the problem, could you please tell me the solution? I would be very grateful. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for your patience and willingness to continue adjusting the parameters. I understand how challenging it can be to fine-tune models for optimal performance. 🛠️ Once I find a solution or any improvement in the training process, especially concerning the hyperparameters that significantly impact the mAP, I'll make sure to share the details with you. It's all about trial and error and learning from each step we take. In the meantime, considering your focus on pose estimation, I recommend revisiting the Ultralytics YOLOv8 Pose Estimation documentation. It provides insights into pre-trained models, training, validation, and prediction processes specifically tailored for pose estimation tasks. Sometimes, a small adjustment or a different approach can lead to significant improvements. Keep experimenting with different configurations, and don't hesitate to reach out if you have more questions or share any progress. Good luck! 🍀 |
Beta Was this translation helpful? Give feedback.
-
|
No problem. We'll share our learning experiences with each other. If you pull it out, please let me know
…---Original---
From: ***@***.***>
Date: Mon, Mar 11, 2024 16:23 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
Thank you for your reply. I will continue to adjust the parameters and hope to make some improvements. After you solve the problem, could you please tell me the solution? I would be very grateful.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
ok，it's my pleasure. |
Beta Was this translation helpful? Give feedback.
-
|
@2823387967 非常乐意分享和交流学习经验！如果我找到了解决方案，一定会告诉你的。同时，如果你对姿势估计任务有任何疑问，可以参考我们的姿势估计文档，里面有关于如何使用YOLOv8进行姿势估计的详细指导，包括预训练模型的使用、如何训练和验证你的模型等内容。希望对你有所帮助！🙂 # 示例：使用预训练的YOLOv8n-pose模型进行姿势估计
from ultralytics import YOLO
# 加载模型
model = YOLO('yolov8n-pose.pt') # 加载官方模型
# 进行预测
results = model('https://ultralytics.com/images/bus.jpg') # 对图像进行预测如果有任何问题，欢迎随时提问！ |
Beta Was this translation helpful? Give feedback.
-
|
I have made some progress by reducing the learning rate to 0.001. After training for 100 epochs, I found that mAP50 has increased by 10%, but I used pre trained weights. |
Beta Was this translation helpful? Give feedback.
-
|
Is there a hyperparameter in YOLOv8 that can reduce the learning rate faster during training |
Beta Was this translation helpful? Give feedback.
-
|
@2823387967 yes, YOLOv8 includes a hyperparameter called yolo detect train data=coco128.yaml model=yolov8n.pt epochs=100 lr0=0.01 lrf=0.1In this command, |
Beta Was this translation helpful? Give feedback.
-
|
How close are you to official right now?Can you show me your hyperparameter Settings?thank you very much！
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 11:33 AM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
Thanks for your reply, I used pre-training weights to reduce the learning rate, and found that the mAP increased by 10% compared to the previous experiment, but after 30 rounds the mAP increased very slowly, from 30 rounds to 100 rounds only increased by 3%.Is there a hyperparameter in YOLOv8 that can reduce the learning rate faster during training.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
Here is my hyperparameter,but I think it's a problem with the learning rate |
Beta Was this translation helpful? Give feedback.
-
|
@2823387967 it looks like you're exploring adjustments to the learning rate for your YOLOv8 pose estimation task. Given your hyperparameters, you might consider experimenting with a learning rate scheduler to dynamically adjust the learning rate during training. For instance, using a Here's a quick adjustment to your hyperparameters to include a cosine learning rate schedule: cos_lr: true
lr0: 0.01 # initial learning rate
lrf: 0.2 # final learning rate multiplier, lr0 * lrfThis setup starts with a higher initial learning rate ( Remember, the effectiveness of hyperparameter adjustments can vary based on the specific dataset and task, so it's always a good idea to experiment and monitor the results closely. Good luck with your training! 🚀 |
Beta Was this translation helpful? Give feedback.
-
|
map50 how much now
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 10:25 AM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
I have made some progress by reducing the learning rate to 0.001, training for 100 epochs, and found that mAP50 has increased by 10%
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
The mAP50 is about 76%, which is 10% lower than the official |
Beta Was this translation helpful? Give feedback.
-
|
It's great to hear about your progress! Achieving a mAP50 of 76% is a solid result, but I understand aiming to match or exceed the official benchmarks is always the goal. Here are a couple of suggestions that might help improve your model's performance further:
Remember, achieving high accuracy is often a process of trial and error, so keep experimenting! 🚀 |
Beta Was this translation helpful? Give feedback.
-
|
Did you load the pre-training weights?
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 12:02 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
The mAP50 is about 76%, which is 10% lower than the official
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
yes,but after 30 rounds the mAP increased very slowly, from 30 rounds to 100 rounds only increased by 3%. |
Beta Was this translation helpful? Give feedback.
-
|
It's quite common for the mAP to increase more slowly after the initial epochs, as the model begins to converge to its optimal state. To potentially boost the mAP increase rate, you might consider adjusting the learning rate schedule, experimenting with different data augmentation strategies, or even extending the training with more epochs if you're not overfitting. Additionally, reviewing the learning rate and loss graphs can provide insights into whether the training is plateauing or if there's still room for improvement. For pose estimation tasks specifically, ensuring your dataset is diverse and representative of the test scenarios can also help improve performance. Since you're working with YOLOv8 for pose estimation, leveraging the # Example command to adjust learning rate and epochs
yolo pose train data=coco8-pose.yaml model=yolov8n-pose.pt epochs=150 lr0=0.005Remember, incremental improvements are still valuable, especially in complex tasks like pose estimation. Keep experimenting! 🚀 |
Beta Was this translation helpful? Give feedback.
-
|
yes，My situation is the same as yours, with lower results，map50 only have 66.8％
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 12:22 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
yes,but after 30 rounds the mAP increased very slowly, from 30 rounds to 100 rounds only increased by 3%.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
Have you ever tried to reduce your learning rate? |
Beta Was this translation helpful? Give feedback.
-
|
@2823387967 确实，降低学习率有时可以帮助提高模型的性能，特别是当你发现模型的mAP提升缓慢或者已经达到一个平台期时。你可以尝试逐步减小学习率，观察是否有所帮助。例如，如果你当前的学习率是0.01，你可以尝试将其减半到0.005，然后继续训练，看看结果是否有所改善。这里是一个简单的代码示例来调整学习率： yolo pose train data=coco8-pose.yaml model=yolov8n-pose.pt epochs=100 imgsz=640 lr0=0.005此外，确保你的训练数据足够多样化且质量高，这也是提高模型性能的关键因素。希望这些建议能帮到你！🙂 |
Beta Was this translation helpful? Give feedback.
-
|
No,is default
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 12:41 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
Have you ever tried to reduce your learning rate?
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
Reducing the learning rate can improve the mAP, I use 0.001, I can reach 75% with 100 rounds of mAP, if the learning rate is 0.0008 or less, it may increase the mAP |
Beta Was this translation helpful? Give feedback.
-
|
Yes, the learning rate is set to default values during training, but you can easily adjust it according to your needs. For instance, if you're training a pose estimation model like yolo pose train data=coco8-pose.yaml model=yolov8n-pose.pt epochs=100 imgsz=640 lr0=0.01This command sets the initial learning rate to 0.01. Adjusting the learning rate can be crucial for achieving better model performance, so feel free to experiment with different values! 😊 |
Beta Was this translation helpful? Give feedback.
-
|
Thank you very much for your reply. I am also making some attempts. Please keep in touch. If you can find out the official results, please let me know.
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 15:37 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
Reducing the learning rate can improve the mAP, I use 0.001, I can reach 75% with 100 rounds of mAP, if the learning rate is 0.0008 or less, it may increase the mAP
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
@AliasChenYi hey there! 🌟 Thanks for reaching out and making efforts on your end. I'll definitely keep you updated with any official results we find. In the meantime, if you have any more questions or need further assistance, feel free to ask. Let's keep the communication open! Happy coding! 😊 |
Beta Was this translation helpful? Give feedback.
-
|
Can you tell me the name of the pre-trained weight file you use?I suspect I used it wrong.
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 15:37 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
Reducing the learning rate can improve the mAP, I use 0.001, I can reach 75% with 100 rounds of mAP, if the learning rate is 0.0008 or less, it may increase the mAP
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
I use yolov8n-pose.pt |
Beta Was this translation helpful? Give feedback.
-
|
Do you have any other contact details? Wechat or TIM? |
Beta Was this translation helpful? Give feedback.
-
|
我们目前主要通过GitHub进行沟通，以确保问题和讨论的透明度和可追溯性。如果您有任何问题或需要进一步的帮助，请随时在相应的GitHub issue下留言，我们的团队会尽快回复您。感谢您的理解和支持！🙌 另外，关于您提到的预训练权重文件， |
Beta Was this translation helpful? Give feedback.
-
|
w：shanghw618
…---Original---
From: ***@***.***>
Date: Tue, Mar 12, 2024 18:24 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
Do you have any other contact details? Wechat or TIM?
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
Hey there! 👋 Unfortunately, we don't share personal contact details like WeChat or TIM publicly for privacy and security reasons. However, you're more than welcome to reach out through the issues section here on GitHub for any questions or discussions related to Ultralytics YOLOv8. We're here to help! If there's something specific you need assistance with, feel free to ask right here. 😊 For pose estimation tasks with YOLOv8, you might find the following snippet useful for starting your training: from ultralytics import YOLO
# Load a pretrained pose estimation model
model = YOLO('yolov8n-pose.pt')
# Train the model on your dataset
results = model.train(data='your_dataset.yaml', epochs=100, imgsz=640)Let us know if you have any more questions or if there's anything else we can assist you with! |
Beta Was this translation helpful? Give feedback.
-
|
For HPE, I am facing a problem on counting the number of repetitive movements in video, how can I develop a HPE algorithm for this purpose? Thank you very much! |
Beta Was this translation helpful? Give feedback.
-
|
Hey there! 👋 For counting repetitive movements in videos using Human Pose Estimation (HPE) with YOLOv8, you can start by using our pretrained pose models like Here's a basic approach:
For example, to load and use a pretrained pose model: from ultralytics import YOLO
# Load a pretrained pose model
model = YOLO('yolov8n-pose.pt')
# Predict keypoints on an image
results = model('path/to/your/image.jpg')For tracking and analyzing movements, you might need to implement custom logic based on the keypoints' movements across frames. Hope this helps! Let us know if you have any more questions. 😊 |
Beta Was this translation helpful? Give feedback.
-
|
how could i directly get the box with target's xyhw from the results of model prediction like "results = model.predict(source=frame)" as some parameters for other functions |
Beta Was this translation helpful? Give feedback.
-
|
Hey there! 😃 To directly extract the bounding box (x, y, width, height) details from your prediction results for further processing, you can simply access the from ultralytics import YOLO
# Load your model
model = YOLO('yolov8n.pt')
# Run prediction
results = model.predict(source=frame)
# Access the bounding box details
for result in results:
boxes = result.boxes.xywh # This gives you a tensor of shape [N, 4] where N is the number of detections
# Now, you can loop over the boxes and use them as needed
for box in boxes:
x, y, w, h = box.cpu().numpy() # Convert to numpy array if needed
print(f"Box coordinates: x={x}, y={y}, width={w}, height={h}")This loop will print the x, y, width, and height for each detection in your frame. You can then use these coordinates as parameters for your other functions. 🚀 Happy coding! |
Beta Was this translation helpful? Give feedback.
-
|
import pygetwindow as gw Load the modelmodel = YOLO("model/best.onnx") while True: I need help, who will help me? |
Beta Was this translation helpful? Give feedback.
-
|
@3300299472 hello! It's great to hear that you've successfully used the ONNX model for object detection. To obtain the coordinates of the detected objects and control the mouse accordingly, you can extract the bounding box coordinates from the from ultralytics import YOLO
import pyautogui
# Load the model
model = YOLO("model/best.onnx")
# Perform inference
results = model.predict(source=img, show=True)
# Get bounding box coordinates
for det in results.xyxy[0]: # detections per image
x1, y1, x2, y2, conf, cls = det
# Calculate the center of the bounding box
x_center = int((x1 + x2) / 2)
y_center = int((y1 + y2) / 2)
# Move mouse to the center of the bounding box
pyautogui.moveTo(x_center, y_center)
# Optionally, perform a click
pyautogui.click()This code snippet will move the mouse to the center of each detected object and perform a click. Make sure to adjust the coordinates based on your specific screen resolution and application window size if necessary. Happy coding! 😊 |
Beta Was this translation helpful? Give feedback.
-
|
Hey, I have a problem, I'm going to write a judgment script and I annotated_frame = results[0].plot() and I get the labeled prediction box so I get results = model(frame) and I print results[0] and I get the result, However, only the coordinates of the prediction box did not find the corresponding label, how should I find the class id (I plan to warn if I judge lying for more than five frames)? boxes: ultralytics.engine.results.Boxes object orig_shape: (544, 960) |
Beta Was this translation helpful? Give feedback.
-
|
@ai741 to find the class ID corresponding to the prediction box, you can access the |
Beta Was this translation helpful? Give feedback.
-
|
Why is YOLOv8-Pose performing poorly on animal pose estimation tasks, and what aspects should be improved?And how to improve it? |
Beta Was this translation helpful? Give feedback.
-
|
@legend-zhen yOLOv8-Pose may underperform on animal pose estimation tasks due to differences between the COCO keypoints dataset (human-centric) and animal anatomy. To improve performance, consider fine-tuning a pretrained model on an animal-specific keypoints dataset. If no such dataset exists, you can create one using tools like JSON2YOLO for formatting. Additionally, leveraging domain-specific augmentation and ensuring balanced keypoint representation in your dataset can help. |
Beta Was this translation helpful? Give feedback.
-
|
Hello, what is the principle of YOLOv8 for pose estimation? I am working on improving the YOLOv8-pose network to achieve better pose estimation results, but I have found that the keypoint detection accuracy seems to be unrelated to the detection box accuracy in my experiments, so I am unable to make effective targeted improvements to the network. How should I modify the network to achieve better results |
Beta Was this translation helpful? Give feedback.
-
|
@legend-zhen the key to improving YOLOv8-pose performance lies in understanding its architecture and the relationship between detection and keypoint estimation. While detection accuracy provides bounding box priors, keypoint estimation relies on the model's ability to localize features within those boxes. To enhance results, consider modifying the backbone or neck for better feature extraction, optimizing the loss function to prioritize keypoints, or using higher-quality datasets. For detailed guidance, the Pose Estimation documentation is a good resource. |
Beta Was this translation helpful? Give feedback.
-
|
Hi,I have a question.YOLOv8 is anchor-free, but YOLOv8-pose seems to define anchor points in the pose estimation task. Is it predicting the keypoint positions based on the offsets relative to the anchor points? |
Beta Was this translation helpful? Give feedback.
-
|
@legend-zhen thank you for your question! YOLOv8-pose does not use anchor points in the traditional sense. Instead, it predicts keypoint positions directly as offsets relative to the bounding box of the detected object. This approach aligns with the anchor-free architecture of YOLOv8. For more details on how YOLO-pose works, you can refer to the Pose Estimation documentation. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for your reply! YOLOv8-pose first predicts the object detection box and then predicts the positions of the keypoints based on the offsets relative to the detected bounding box. Is my understanding correct?
…------------------ 原始邮件 ------------------
发件人: "Glenn ***@***.***>;
发送时间: 2024年12月18日(星期三) 晚上11:18
收件人: ***@***.***>;
抄送: ***@***.***>; ***@***.***>;
主题: Re: [ultralytics/ultralytics] zh/tasks/pose/ (Discussion #8788)
@legend-zhen thank you for your question! YOLOv8-pose does not use anchor points in the traditional sense. Instead, it predicts keypoint positions directly as offsets relative to the bounding box of the detected object. This approach aligns with the anchor-free architecture of YOLOv8. For more details on how YOLO-pose works, you can refer to the Pose Estimation documentation.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
@legend-zhen your understanding is correct! YOLOv8-pose predicts keypoint positions as offsets relative to the detected bounding box, aligning with its anchor-free architecture. For further reading, check out the Pose Estimation documentation. |
Beta Was this translation helpful? Give feedback.
-
|
Hello, thank you for the excellent project you provided. Now I want to realize the corner detection of the card. The conventional card can be solved well by using this pose model, but there are a few difficult cases where a corner is located outside the picture. How can I infer the coordinates of the outer corner points from the three corner points already in the picture? |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for your question! Inferring the coordinates of an off-screen corner from three visible points is a geometric problem that can be solved by assuming the object is rectangular and using the relationship between the visible corners. You could calculate the missing corner using vector mathematics or extrapolation based on the known coordinates. For further refinement, integrating a custom logic layer into your YOLO model's post-processing pipeline is recommended. Let us know if you have additional questions! |
Beta Was this translation helpful? Give feedback.
-
|
Hi，PROFESSOR！In the keypoint similarity loss function (OKS) of YOLOv8-pose, the scale factor of the keypoints and the sigma function representing the weight of each keypoint, do they remain fixed or do they change according to the dataset in the pose estimation task?thank you！ |
Beta Was this translation helpful? Give feedback.
-
|
@legend-zhen thank you for your question! The scale factor and sigma function in the OKS loss for YOLOv8-pose remain fixed and are predefined depending on the dataset, such as COCO keypoints. If using a custom dataset, these parameters can be adjusted accordingly in the dataset configuration. For more details, you can refer to the COCO-pose configuration. |
Beta Was this translation helpful? Give feedback.
-
zh/tasks/pose/
了解如何使用Ultralytics YOLOv8 完成姿势估计任务。查找预训练模型，学习如何训练、验证、预测和输出自己的模型。
https://docs.ultralytics.com/zh/tasks/pose/
Beta Was this translation helpful? Give feedback.
All reactions