tsize can't input high resolution

[zxq309]

I use yolov5 training input image size is (1600，1600), when inferencing, I want to set tsize to (1600，1600), but there is an error, what happened? tsize = （640,640）is running and correct.

[tungdop2]

For my understanding, this is the input size for the model.

[zxq309]

errors still occur. The input image size is 1600 when I custom trained custom model. change Input size on yolov5 source code is OK. But in your code only [640,640] is OK, [608, 1088], [800, 1440] all occur errors, How to solve it

[tungdop2]

I'm not sure, give me your model and 1 image, I will check later.

[tungdop2]

errors still occur. The input image size is 1600 when I custom trained custom model. change Input size on yolov5 source code is OK. But in your code only [640,640] is OK, [608, 1088], [800, 1440] all occur errors, How to solve it

Try (896, 1600)?

[zxq309]

errors still occur. The input image size is 1600 when I custom trained custom model. change Input size on yolov5 source code is OK. But in your code only [640,640] is OK, [608, 1088], [800, 1440] all occur errors, How to solve it

Try (896, 1600)?

Thank you for your timely reply. [896, 1600] is correct, everything is OK.
But Input size [640, 1152] can inference but drawn boxes shift, Input size [800, 1440] occurs "RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 26 but got size 25 for tensor number 1 in the list."
How to explain this?

[zxq309]

As your have mentioned，Input size [608, 1088], [800, 1088], [896,1440] all invalid and occur errors when inference videos (2.mp4).
url：https://pan.baidu.com/s/14RJ8vCBfm5X4BwM6AZvrIw
code：njb5
This is the test videos and model, look forward to your reply.

[tungdop2]

I get the numbers from the original repo. For my understanding, with [608, 1088], 608 / image width < 1088 / image height, the model only cares about 608, and that is not the true scale for yolov5x6, then the boxes shift. With [896,1440], the model gets 1440, and this is not a valid input. You can get it at update function in tracker/bytetracker.py.

[zxq309]

So ，How can I change this. Arbitrary image width and height can be entered

[tungdop2]

No, I think it must be 1 of that numbers.

[zxq309]

I change different input images, but errors occur once more. Can the author change this status, or reply more clearly? boxes drift is too serious！

[tungdop2]

It's based on your model input size, not your image size.

[zxq309]

OK, I see that's model input size. I set model input size [832,1536], video frame size is [4000,3000], but boxes drift is too serious, what's wrong? and how to change, I need chage image size to [832, 1536]?