When inference spec and input image dimensions differ

Hello,

I have noticed that the tao inference command still runs even if the image_width and image_height in the inference spec do not match the dimensions of the input image(s). Specifically I am running detectnet_v2 or one of its variants.

My question is, what is the inference command actually doing in this scenario? Is it resizing the input image, cropping, something else?

Refer to DetectNet_v2 - NVIDIA Docs

Note
Since detectnet is a full convolutional neural net, the model can be inferred at a different inference resolution than the resolution at which it was trained. The input dims of the network will be overridden to run inference at this resolution, if they are different from the training resolution. There may be some regression in accuracy when running inference at a different resolution since the convolutional kernels don’t see the object features at this shape.

Hi Morgan,

Thanks for the prompt response. However, I don’t think that clearly addresses my question.

My question is not about whether it’s ok for training dimensions to not match inference dimensions, nor am I asking whether inference still occurs when the spec differs from the input image.

I am asking what actually happens when the image dims stated in the inference spec differs from the image dims at inference time.

So when the doc says “The input dims of the network will be overridden to run inference at this resolution,” it isn’t clear which resolution that refers to. Does “this resolution” refer to the dims in the inference spec, or the dims of the input image?

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

“this resolution” refer to the dims in the inference spec.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.