Hello, when I use my own depth map and rgb (h, w is 900x1600) for inference, I put the "--convolutional-layer-" The encoding" parameter is changed to std, there is a problem that the model does not match the weight in the backbone part when loading the weight. but I see that when you are training, the "--convolutional-layer-encoding" parameter defaults to xyz, so should I train a model with the "--convolutional-layer-encoding" is "std" from scratch? Besides, I wonder if changing this parameter will affect the final performance of the model?
Hello, when I use my own depth map and rgb (h, w is 900x1600) for inference, I put the "--convolutional-layer-" The encoding" parameter is changed to std, there is a problem that the model does not match the weight in the backbone part when loading the weight. but I see that when you are training, the "--convolutional-layer-encoding" parameter defaults to xyz, so should I train a model with the "--convolutional-layer-encoding" is "std" from scratch? Besides, I wonder if changing this parameter will affect the final performance of the model?