Question about pose file format and external_cond_dim for custom dataset adaptation

[fangchuan]

Hi, thanks for your awsome work. I'm currently adapting the codebase to work with my custom dataset and have a question about the pose file format and configuration.

I noticed that the repository requires cond_pose files, and when inspecting the existing *.pt files in test_poses and training_poses directories, I found they are tensors of shape (num_frames, 18):

cond = torch.load(test_pose_file)
print(cond.shape)  # torch.Size([174, 18])

However, looking at the configuration in configurations/dataset/realestate10k.yaml, the external_cond_dim parameter is set to 16 (as shown in the screenshot below):

This discrepancy raises a couple of questions:

• Preprocessing: Could you explain how the *_poses files are preprocessed? What does each dimension in the 18-dimensional tensor represent?

• Camera parameter injection: How are these pose parameters mapped to the external_cond_dim=16 expected by the backbone? I'd like to understand the pipeline to properly prepare pose files for my custom dataset.

Any guidance or documentation on this would be greatly appreciated! @kwsong0113 @buoyancy99