![]() When it comes to annotations, YOLOv8 uses relative coordinates rather than absolute pixel values for the bounding box positions. In your case, with the original input image of 1920x1080, the images will be resized to an aspect ratio close to 640x360. When providing image size (imgsz=640) for training your model, YOLOv8 takes care of resizing the images to have their longest dimension set to 640 while maintaining the image's original aspect ratio. I guess this can also help with training. It was mentioned to use rect if image is not square. Some tutorials on the internet can be misleading, that this why I want to understand the input for the model. This is my first time using YOLO and I don't have experience with previous versions. For example, if I take model=YOLOv8x-seg can I pass imgsz=3840 ? I understand that it might take longer and consume more memory. The class (power lines) I'm trying to segment is very small in the image, I want to use the original size as much as I can. Q3: If I'm using already pretrained models/weights (in my case YOLOv8x-seg), the table says the size is 640, can I use it for training on custom size. Does it mean we can provide either one size (longest side size of the image) or both sizes as w, h? Can we do imgsz=3840,2160 for example? Q2: For imgsz parameter, documentation says "size of input images as integer or w,h". In this case if I feed images with size of 3840x2160, the model will convert them to 640x320, right? It was also mentioned somewhere that the model will keep the w and h ratio. Q1: When we are providing imgsz=640 does the model resize the original size to 640? (I believe it was mentioned in the repo). I'm working on semantic segmentation problem and have a custom dataset with the image size of 3840x2160. I checked the documentation and the repo but couldn't find related information to this. I have searched the YOLOv8 issues and discussions and found no similar questions.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |