I used yolov11n and yolov11x to train many models based on this dataset. The log accuracy of small batch and large epoch training is very good, but it is very poor during actual inference, much worse than mediaPipa, especially when rotating, the key points are very inaccurate. I am very confused about this, is there any trick to the training parameters?
It sounds like your model might be struggling with generalization during inference, especially for rotations and keypoint accuracy. Ensure the following: 1. Data Augmentation: Use robust augmentations like rotation and scaling to improve the model's ability to generalize to rotated objects. Learn more about augmentation here: Data Augmentation docs.ultralytics.com/reference/data/augment/. 2. Class Distribution and Label Quality: Verify that the dataset's annotations and class distribution are balanced and accurate. Poor annotations can lead to suboptimal learning. Refer to Dataset Format and Labels docs.ultralytics.com/guides/yolo-common-issues/. 3. Pretrained Weights: Fine-tuning a pre-trained YOLO11 model rather than training from scratch may yield better results, especially for challenging tasks like keypoint detection. More details here: Pretrained Weights docs.ultralytics.com/guides/yolo-common-issues/. 4. Hyperparameter Tuning: Adjust learning rates, batch sizes, and epochs. Small batch sizes with high learning rates can sometimes cause issues with convergence. For guidance, check Hyperparameter Tuning docs.ultralytics.com/reference/utils/tuner/. 5. Oriented Bounding Box (OBB): For rotations, training with OBB models might improve precision. Learn more: OBB Models docs.ultralytics.com/tasks/obb/. If the issue persists, try visualizing the prediction outputs to diagnose specific errors in keypoint detection.
Thanks for pointing that out! Even a small percentage of incorrect labels, like 5%, can significantly impact training quality and inference accuracy, especially for tasks like pose estimation. To mitigate this, consider cleaning the dataset by reviewing and correcting annotations or applying noise-tolerant training techniques. You can also incorporate tools to visualize and verify labels during data preprocessing. Here's a guide on dataset preparation and annotation: Data Collection and Annotation docs.ultralytics.com/guides/data-collection-and-annotation/.
@@Ultralytics I have cleaned the data. In fact, 15% of the data is incorrectly labeled. After training on the cleaned data set, the generalization ability is significantly improved. Do you need me to clean the data set?
Great to hear that cleaning the dataset improved generalization! A well-labeled dataset is crucial for optimal model performance, as incorrect labels can lead to poor learning and predictions. If 15% of the data was mislabeled, cleaning it definitely made a significant difference. While we appreciate your offer, we encourage users to clean datasets tailored to their specific needs. You can find helpful guidelines on data annotation and verification here: Data Collection and Annotation docs.ultralytics.com/guides/data-collection-and-annotation/. Thank you for contributing to better results! 😊
Great question! The model primarily detects keypoints, such as hand joints, from the dataset. Using these keypoints, you can infer gestures by applying logic or additional algorithms. The dataset provides detailed annotations for 21 hand keypoints, but gesture classification, if needed, would require additional training or post-processing. Learn more about the dataset and keypoints here docs.ultralytics.com/datasets/pose/hand-keypoints/. 😊
Thanks for the feedback! If you're encountering corrupt image warnings, try checking your dataset for any issues or re-download it if possible. For the Weights & Biases API, ensure you've set it up correctly in your environment. You can find more setup details here: docs.ultralytics.com/integrations/weights-biases/. Let us know if you need further assistance! 😊
@Ultralytics I don't have any issues. But if you look through this video, it getting warnings and.... I think the guy in the video doesn't care about validity of running codes. tnx btw
Thanks for pointing that out! We strive to ensure our tutorials are as accurate and helpful as possible. I'll pass your feedback on to the team to improve future content. If you have any other suggestions or questions, feel free to share. 😊
Thanks for sharing this amazing tutorial
You're welcome! I'm glad you enjoyed it! If you have any questions or need further information, feel free to ask. 😊
I used yolov11n and yolov11x to train many models based on this dataset. The log accuracy of small batch and large epoch training is very good, but it is very poor during actual inference, much worse than mediaPipa, especially when rotating, the key points are very inaccurate. I am very confused about this, is there any trick to the training parameters?
It sounds like your model might be struggling with generalization during inference, especially for rotations and keypoint accuracy. Ensure the following:
1. Data Augmentation: Use robust augmentations like rotation and scaling to improve the model's ability to generalize to rotated objects. Learn more about augmentation here: Data Augmentation docs.ultralytics.com/reference/data/augment/.
2. Class Distribution and Label Quality: Verify that the dataset's annotations and class distribution are balanced and accurate. Poor annotations can lead to suboptimal learning. Refer to Dataset Format and Labels docs.ultralytics.com/guides/yolo-common-issues/.
3. Pretrained Weights: Fine-tuning a pre-trained YOLO11 model rather than training from scratch may yield better results, especially for challenging tasks like keypoint detection. More details here: Pretrained Weights docs.ultralytics.com/guides/yolo-common-issues/.
4. Hyperparameter Tuning: Adjust learning rates, batch sizes, and epochs. Small batch sizes with high learning rates can sometimes cause issues with convergence. For guidance, check Hyperparameter Tuning docs.ultralytics.com/reference/utils/tuner/.
5. Oriented Bounding Box (OBB): For rotations, training with OBB models might improve precision. Learn more: OBB Models docs.ultralytics.com/tasks/obb/.
If the issue persists, try visualizing the prediction outputs to diagnose specific errors in keypoint detection.
@@Ultralytics In fact, the official dataset has 5% incorrect labels.
Thanks for pointing that out! Even a small percentage of incorrect labels, like 5%, can significantly impact training quality and inference accuracy, especially for tasks like pose estimation. To mitigate this, consider cleaning the dataset by reviewing and correcting annotations or applying noise-tolerant training techniques. You can also incorporate tools to visualize and verify labels during data preprocessing. Here's a guide on dataset preparation and annotation: Data Collection and Annotation docs.ultralytics.com/guides/data-collection-and-annotation/.
@@Ultralytics I have cleaned the data. In fact, 15% of the data is incorrectly labeled. After training on the cleaned data set, the generalization ability is significantly improved. Do you need me to clean the data set?
Great to hear that cleaning the dataset improved generalization! A well-labeled dataset is crucial for optimal model performance, as incorrect labels can lead to poor learning and predictions. If 15% of the data was mislabeled, cleaning it definitely made a significant difference. While we appreciate your offer, we encourage users to clean datasets tailored to their specific needs. You can find helpful guidelines on data annotation and verification here: Data Collection and Annotation docs.ultralytics.com/guides/data-collection-and-annotation/. Thank you for contributing to better results! 😊
I have question, is this model just detecting the keypoints from the dataset or it is detecting the gesture?
Great question! The model primarily detects keypoints, such as hand joints, from the dataset. Using these keypoints, you can infer gestures by applying logic or additional algorithms. The dataset provides detailed annotations for 21 hand keypoints, but gesture classification, if needed, would require additional training or post-processing. Learn more about the dataset and keypoints here docs.ultralytics.com/datasets/pose/hand-keypoints/. 😊
You got so many warnings for corrupt images. Also didn't put wandb api. WEAK
Thanks for the feedback! If you're encountering corrupt image warnings, try checking your dataset for any issues or re-download it if possible. For the Weights & Biases API, ensure you've set it up correctly in your environment. You can find more setup details here: docs.ultralytics.com/integrations/weights-biases/. Let us know if you need further assistance! 😊
@Ultralytics I don't have any issues. But if you look through this video, it getting warnings and.... I think the guy in the video doesn't care about validity of running codes. tnx btw
Thanks for pointing that out! We strive to ensure our tutorials are as accurate and helpful as possible. I'll pass your feedback on to the team to improve future content. If you have any other suggestions or questions, feel free to share. 😊