Normalize image before inference

[mihelich]

Thanks for the helpful repo!

When using the pretrained models, should I normalize the image tensor before feeding it to the model?

For training, COCODataset applies the usual Normalize transform (code) when loading images. The config files also include NormalizeTensor steps in train_pipeline and val_pipeline, though that part of the config doesn't appear to be used in this repo.

On the inference side, however, the inference.py script only scales the image input to [0, 1], without normalizing (code). Is that a bug, or am I missing something?

[JunkyByte]

Hello, thanks for your interest and for the bug report, I agree the preprocessing is different in inference, the original implementation I built on (https://github.com/jaehyunnn/ViTPose_pytorch) seems to contain this bug and I did not notice.

Should be fixed with c8a7a14, I inspected the inference results visually and they improve. I referenced this issue on the original implementation.

[JunkyByte]

Feel free to close once you verify this fix works correctly for you as well!

[mihelich]

Great, thanks for the quick response! The fix looks correct to me. I'm using the ViTPose model directly, so I'm happy to have confirmation on what preprocessing I should do.