MessyKitchens cameras
cameras.json contains the final fitted cameras for all 1,503 prepared views.
The schema is partprobe3d.messykitchens.cameras.v1.
The views dictionary uses the same view keys as the prepared dataset manifest.
| Field | Meaning |
|---|---|
scene |
Original scene identifier |
source_view |
Original photograph identifier |
native_size |
Image width and height in pixels, in that order |
projection_intrinsics |
The 3 × 3 intrinsic matrix for the native image |
camera_to_world |
The 4 × 4 rigid transform from camera coordinates to world coordinates |
World coordinates match the prepared GT meshes in meters.
Camera axes point right, down, and forward for positive x, y, and z.
Native pixel centers use integer coordinates, starting at (0, 0).
The intrinsic matrix uses these native pixel coordinates.
For column vectors, camera_to_world maps [x, y, z, 1] from the camera frame to the world frame.
Its inverse provides the world-to-camera transform.
The top-left 3 × 3 block contains rotation.
The final column contains translation in meters.
These cameras result from GT-assisted silhouette fitting and constitute estimated calibration. All views passed local convergence and native overlap plateau checks. The file preserves the final numeric values without rounding. Fit diagnostics, candidate cameras, and optimizer records remain outside this clean export.
This file supplements the existing evaluation archives. Their per-view registration records retain the earlier camera estimates.