Open Leo-aetech opened 2 weeks ago
👋 Hello @Leo-aetech, thank you for your interest in Ultralytics 🚀! We appreciate you highlighting your question and providing details about your recycling classification project 🤖♻️. We recommend taking a look at the Docs for insights on different YOLO tasks such as Detection and Segmentation, as they may help clarify some of the performance discrepancies you're observing.
If this is a ❓ Question about fine-tuning or pre-training, please provide as much additional context as possible. Specifically:
If you're facing a 🐛 Bug related to training or performance, please share a minimum reproducible example (MRE) to help us reproduce and debug the issue effectively.
For your workflow query, training strategies like pre-training and fine-tuning are indeed possible, and the Model Training Tips guide might provide helpful insights to enhance your results. However, remember that segmentation and detection models have inherently different objective optimizations and performance benchmarks.
Before diving deeper, ensure you're using the latest version of ultralytics
with all its requirements in a Python>=3.8 environment with PyTorch>=1.8. Updating ensures any resolved issues in newer versions are accounted for:
pip install -U ultralytics
YOLO features can be run in the following environments, allowing you to experiment with different settings:
Our Ultralytics community is an excellent resource to learn and share:
This badge signifies if Ultralytics CI tests are passing. Status reflects verification across all YOLO Modes and Tasks in macOS, Windows, and Ubuntu environments.
This is an automated response to assist you efficiently 🤖. An Ultralytics engineer will follow up soon to better address your question or concerns.
You can pass the path to the trained detector using the pretrained
argument.
Search before asking
Question
I am developing an AI for automatic recycling classification. When using YOLOv11's Detector, I achieve around 85% for mAP 0.5-1. However, when training Segmentation with the same data, it only reaches 50%. Why does this phenomenon occur? Is it possible to pre-train the Detector and then fine-tune Segmentation? If possible, would this help improve performance?
Additional
No response