The combination of segmentation and pose with strange result

[Ichiruchan]

Hello everyone,

I’m currently facing a challenge with my model, where I’ve combined the segmentation head and pose head into a single structure. I’ve adjusted the data reading process and modified the loss function to train the new model with the default hyperparameters. However, the predictions seem off, and the metrics are not performing well (MAP50-95 is about 0.91). For instance, the keypoints are appearing outside the bounding boxes, and both the segmentation and detection components are underperforming

Interestingly, when I remove the keypoint annotations and train on segmentation, the model performs well (MAP50-95 is nearly 0.955).

Could anyone provide suggestions on how to improve this situation?

[UltralyticsAssistant]

👋 Hello @Ichiruchan, thank you for sharing your detailed query and interest in the Ultralytics repository 🚀! Your results and observations are quite intriguing.

We recommend referring to the Docs for detailed guidance on custom setups like this. For additional context, you can review the sections on model configurations and custom training to ensure all modifications align with best practices.

If this is a 🐛 Bug Report related to the model behavior, please share a minimum reproducible example (MRE). For example, please include the following:

1. A concise Python code snippet that replicates this behavior.
2. A description of your data format and annotations, along with a few example images.
3. Logs of the training process, including loss behavior and any irregularities observed.

If this is a ❓ Question about custom training adjustments, please provide:

1. Details of the changes made to the architecture (combination of segmentation and pose head).
2. Adjustments to the data loading process and loss function.
3. Details about your data (keypoint and segmentation annotations, dataset size, classes, etc.).

Upgrade

It's also a good idea to ensure your environment is up to date. You can upgrade to the latest ultralytics package in a Python>=3.8 environment with the following:

pip install -U ultralytics

Environments

Using a verified and stable environment can help debug more reliably. Supported environments include:

• Notebooks:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image: See Docker Quickstart Guide

Community and Resources

To connect with the community and gain further insights, consider joining:

• Discord 🎧 for real-time discussions.
• Discourse for more detailed exchanges.
• Subreddit for tips, tricks, and shared experiences with others.

Status

Lastly, you can check the current CI status of the repository here: . A green badge ensures our continuous integration tests are passing for all Modes and Tasks.

This is an automated response 😊, but don’t worry—an Ultralytics engineer will take a closer look at your issue and assist you shortly!

[glenn-jocher]

@Ichiruchan for multi-task learning combining segmentation and pose estimation, we recommend: 1) Verifying your annotation alignment between segmentation masks and keypoints 2) Adjusting loss weights specifically for pose estimation using the hyp.pose parameter in your hyperparameters file 3) Considering separate task-specific models as combined architectures can create conflicting gradients. The Pose Estimation Documentation shows proper implementation for individual tasks. For custom architectures, you might need to modify the loss balancing - see our pose loss implementation in ultralytics/utils/loss.py.

[Ichiruchan]

I have confirmed that the annotation alignment is correct.
Even after setting the hyperparameters for pose and kobj to 0, there was no improvement for seg.
My target is that each object has only one mask and its corresponding keypoints.
And the plot of labels is correct.

[glenn-jocher]

@Ichiruchan thank you for sharing these details. For combined segmentation and pose estimation, we recommend: 1) Ensuring your dataset format matches the YOLO multi-task requirements shown in the Dataset Documentation, 2) Trying separate task heads rather than combined outputs to avoid feature conflict, and 3) Verifying your loss function properly handles both mask and keypoint supervision simultaneously. The segmentation and pose losses are designed as separate heads in our implementation - you might need to adapt our v8PoseLoss and v8SegmentationLoss from ultralytics/utils/loss.py for combined use.