Transform conversion from colmap `transform.json` to `camera_path.json`

[hyqshr]

For the same frame in a dataset:

Result from transforms_train.json (ns-export camera)

        {
            "file_path": "images/frame_00001.jpg",
            "transform_matrix": [
                [
                    0.9837186500788239,
                    -4.931517566535353e-07,
                    0.17971537910499882,
                    -3.0586481911644254
                ],
                [
                    0.17957862323929175,
                    -0.039001623760136724,
                    -0.9829701884693925,
                    -6.690857223027059
                ],
                [
                    0.0070096763532386825,
                    0.999239147223441,
                    -0.03836653611669037,
                    -0.8800965851079441
                ],
                [
                    0.0,
                    0.0,
                    0.0,
                    1.0
                ]
            ],
            "colmap_im_id": 1
        },

Result from colmap transform.json:

    {
        "file_path": "processed/ARRecording_DCF861EF-2ACC-4834-81FF-11E558FC4F83/images_2/frame_00001.jpg",
        "transform": [
            [
                0.9834007024765015,
                -0.03150152042508125,
                0.17869171500205994,
                -0.48375824093818665
            ],
            [
                0.18097257614135742,
                0.09910257160663605,
                -0.9784823060035706,
                -1.0
            ],
            [
                0.013114869594573975,
                0.99457848072052,
                0.10315844416618347,
                -0.0028430684469640255
            ]
        ]
    },

Do transforms_train.json (from ns-export camera) share the same coordinates with camera_path.json from splatfacto? What's the conversion from colmap to camera_path.json coordinates? Is there any scale difference?

[lyg1597]

My understanding is there's an additional layer of transformation between pose in transform.json and pose actually used by nerfstudio while training. That transformation is applied in nerfstudio_dataparser.py:L236-249. In my understanding these lines does following to the original colmap camera pose.

    camera_pose = transform@camera_pose
    camera_pose[:3,3] *= scale_factor

The transformation matrix and scale factor can be found in dataparser_transforms.json in the output folder. The applied_transform field in transform.json is not used I think.