Please provide steps for custom dataset, also how to generate the transform.json file about zero123

[zz7379]

same waiting this

[lirus7]

Hey @ruoshiliu , Is the transform.json just the blender rotation matrix or is it some camera pose?

[LiPeterzette]

Based on my understanding of the format of this json file, the transform_matrix should be the camera extrinsic matrix in the world coordinate system, representing the transformation from the world coordinate system to the camera local coordinate system, rather than the rotation matrix in Blender. Specifically:

1. The matrix is 4x4, which conforms to the size of the camera extrinsic matrix.

2. The first 3 row vectors represent the directions of the three axes in the camera local coordinate system under the world coordinate system.

3. The 4th row represents the position of the origin of the camera local coordinate system under the world coordinate system.

4. Different transform_matrices reflect different camera perspectives.

These are all typical structures and semantics of the camera extrinsic matrix representing camera pose.

5. While in Blender, the rotation matrix is usually 3x3, representing only three-dimensional rotation transformations.

So I think it can be determined that the transform_matrix in this json represents the camera extrinsic matrix in the NeRF data format, rather than the rotation matrix in Blender.

If it is as I guess, what puzzles me is that the link "file_path": "./train/r_0" is the same for different camera angles given so far, so how to ensure the 3D reconstruction effect? If they're all the same, what does the change in the value of the transform_matrix in different json files determine? Is it just the size of the object? Or is it something else? It is also very likely that there is something wrong with my understanding. I look forward to getting correction and reply