image size input

[andreazuna89]

Hi, thanks for the great work!

I am trying to test your TensorRT script in my application, and I have a question. I noticed some differences between the demo scripts (run_demo and run_demo_tensorrt). In the run_demo_tensorrt script, it seems that the required input image size is [448, 640] or at least "height and width must be divisible by 32".

What should I do if my input images have a resolution of [400, 640]? In the run_demo script I see that image padding is applied, but this part does not appear in the run_demo_tensorrt script. Would it be possible to pad the images before inference and then remove the padding after computing the disparity?

Could you clarify what the recommended input image dimensions are for the best results?

Thanks a lot!

[wenbowen123]

The TRT also neeeds the input size to be divisible by 32. The reason we didn't highlight this in run_demo_tensorrt.py is simply because when we make the onnx files the input size has been set to that. For your case [400, 640], you need to do the maths and specify that image size when you create onnx files. And pad input image during inference if needed.

[andreazuna89]

I understand, thank you very much. In your experience, is it better to create an ONNX model with fixed image dimensions of [400, 640], or to create a model where the height and width are divisible by 32, padding the image during inference and removing the padding after computing the disparity? Would the results be almost the same?

[wenbowen123]

You have to do the latter (the former won't work).