Hi,
I need to be able to detect objects, people, and animals. Ideally, I would use the “Detect” task, but is there any advantage to using “Segment”? I understand that segmentation will still identify the object, but does it provide any additional information or data?
It seems to highlight the object more precisely compared to a simple bounding box around it.
Depends on what you want. Detection gives you boxes. Segmentation also gives you masks. Segmentation is slower than detection. If you need to only get boxes, segmentation would be sacrificing performance for no good reason.
However, Ultralytics lets you train a detection model with a segmentation dataset. You can get slightly higher accuracy if you use a segmentation dataset to train a detection model because the precise labels allow for better quality augmentations.
Thanks, I understand that segmentation would be better, as it allows me to determine the exact location within the frame where the detection is occurring, rather than just using a bounding box.
For example, I would like to define areas within the frame that are not relevant, such as ignoring detections in certain regions. With segmentation, I can more accurately identify whether a person or animal is in a specific area, such as the top-left section.
Yes, for training purposes, segmentation also makes sense, as it enables precise identification of the exact object.
Yep, that’s exactly the main benefit.
With Ultralytics YOLO26 segment models you still get the normal detections (boxes, classes, conf), but you also get a per-object mask, which gives you the true object shape instead of a rectangle. That makes region-based logic much more accurate, especially when an object only partially enters an area or crosses a boundary. The difference is well described in the image segmentation overview and the instance segmentation guide.
So for your “ignore part of the frame” use case, segmentation is better if you want to check actual mask overlap with a zone. If you only need a rough answer like “is the person somewhere near the top-left,” plain detect is usually enough and faster.
In short: if boundaries matter, use segment; if boxes are enough, use detect.
Segment will be better as it will provide more information, plus I can specify what I am expecting to detect, and it will trigger only for that. Is there any documentation I can read to see what else is natively avilable in the Yolo26 library so my code does not replicate any functions that are already native?
But when running in standard detection, on Raspberry Pi, what would be expected FPS? Am I running yoloe-26s-seg and yoloe-26s?
I also have Hailo-8, but would need to port it across to it.