onnx-community / metric3d-vit-XXXX models giving different answers than original pytorch model

#1
by lidarguy - opened

with onnxruntime CPUExecutionprovider (and proper scaling of the image depths accounting for image scaling and canonical intrinsic scaling), the onnx models give different answers on CPU.

is anyone else experiencing this issue?

my assumption is that these exports are metric3dv2 and not v1, @Xenova can you confirm?

also, can you confirm whether the preprocess size for the large and giant models is supposed to be 518,518 as it is in the small model, or should it be the 616,1064 as in the normal model? (the preprocess configs look like they were copied from the small without modification)

Sign up or log in to comment