Why are width and height values in inferencing output switched? about ultralytics

[glenn-jocher]

@haimat hello! The output you're seeing, where it appears that height and width are switched (listed as 1216x1920 instead of 1920x1216), is due to the format typically used in deep learning where dimensions are often described as height x width rather than the conventional width x height. This format is standard in many frameworks and tools that deal with image processing including PyTorch and Ultralytics YOLO.

Regarding your question about using the rect training parameter: using rect can be beneficial as it trains the model using rectangular images rather than square. This can lead to better memory utilization and potentially faster training times, especially if your input images are naturally rectangular (like yours). It adjusts the network to more effectively handle the aspect ratios of your input images during training, which can improve detection accuracy on similarly shaped test images.

Hope this clarifies your concerns! 😊

[haimat]

@glenn-jocher Thanks your response.

So in my case, if I would use the rect parameter, would that also mean
I should provide both width and height for the imgsz parameter?

[glenn-jocher]

Hello @haimat! Yes, using the rect parameter allows the model to train on non-square images, which can be beneficial for handling different aspect ratios more effectively. When using rect, it's a good idea to specify both width and height in the imgsz parameter to match your typical image dimensions, like so: imgsz=1920,1200. This setup helps the model to better adapt to the dimensions of your input images during training. 😊