How to Integrate Ultralytics YOLO Models with FME — Best Practices Guidance

Hi all,

I'm trying to bring object detection capabilities into an FME workflow using YOLO models built with Ultralytics, and I'd love some guidance from anyone who's tackled this before.

What I'm trying to do:

• Run inference with a trained Ultralytics YOLO model (e.g., detection, segmentation, or OBB) inside an FME workspace
•   Data (e.g., from PythonCaller ) into the model and get bounding boxes / masks / class labels back as FME features
• Ideally write the results back out as vector features (polygons/points) with attributes for class, confidence score, etc.

Specific questions:

1. Has anyone integrated Ultralytics' Python package (pip install ultralytics) directly via FME's PythonCaller transformer? Any gotchas with FME's bundled Python environment vs. installing ultralytics and its dependencies (torch, etc.)?
2. Is it better to run YOLO inference as an external subprocess/script call from FME, or embed the from ultralytics import YOLO calls directly inside a PythonCaller?

Sample datasets/models for testing (for reference):

For anyone who wants to test a workflow before working with production data, Ultralytics provides several small, ready-to-use datasets and pretrained models:

• COCO8 (tiny 8-image subset of COCO, great for quick pipeline tests) — 

https://docs.ultralytics.com/datasets/detect/coco8#introduction

• COCO8-seg (segmentation version) — 

https://docs.ultralytics.com/datasets/segment/coco8-seg#introduction

• Full dataset catalog (detection, segmentation, pose, classification, OBB) — 

https://docs.ultralytics.com/models#contributing-new-models

• Pretrained model list (YOLO26, YOLO11, YOLOv8, etc., COCO/ImageNet-pretrained .pt weights) —

https://github.com/ultralytics/ultralytics