Related Formats and Exporters

How the kinds compare with community formats

The geometry kinds were designed against the formats the community already uses. They keep the spirit of those formats but follow this package's rules: one JSON schema per kind, declared conventions, typed frames, and Uncertain values stored with a covariance. The exact foreign syntax is produced only by the exporters. The references below were read from shallow clones taken on 2026-10-09.

KindRelated formatsAdoptedNot adopted
transforms/v5.0ROS geometry_msgs/TransformStamped (common_msgs/geometry_msgs/msg/TransformStamped.msg:1-2,8-10: header.frame_id → child_frame_id); Kalibr camchain T_cn_cnm1, T_cam_imu (kalibr_common/ConfigReader.py:678-717, "LastCamToHere", "ImuToCam")both ends of a transform named in each record; what the transform does to a point, statedthe direction encoded in a variable name (T_to_from); ROS's x,y,z,w quaternion order (Quaternion.msg:3-6), where ours is w,x,y,z and declared
trajectory/v5.0TUM RGB-D t tx ty tz qx qy qz qw (evo file_interface.py:121-174); EuRoC/ASL state_groundtruth_estimate0/data.csv (:224-243, ns, w first, p_RS_R); KITTI odometry 3×4 rows without timestamps (:177-220); evo stores w,x,y,z (core/trajectory.py:64,71)an explicit time per pose (time, in seconds since a declared time_origin); the body → world direction; an untracked frame absent rather than NaNa positional join as in KITTI (frames carry their frame_index and the image name); the unit and the frames are declared, not implied
target_poses/v3.0BOP scene_gt.json (bop_toolkit_lib/inout.py:267-300; docs/bop_datasets_format.md:108-112: per image [{cam_R_m2c, cam_t_m2c, obj_id}]) and scene_camera.json (:79-90, inout.py:221-260)the per-image object list; the model → camera direction; an undetected object absentan integer obj_id (targets keep their identity); a K per image (a frame names its camera record through sensor)
observations/v3.0, 3D pointsOpenSim TRC (OpenSim/Common/TRCFileAdapter.cpp:236-376, reader :38-233: Frame# Time plus X/Y/Z per marker, empty means missing); C3D (ezc3d/README.md:526-612: POINT data, a residual < 0 means occluded)a time per frame; a unit; named markersNaN or empty fields (an unseen marker is absent); the scalar residual (a covariance is stored when it is known)
pose covariance (every kind)ROS PoseWithCovariance (PoseWithCovariance.msg:5-9: row-major 6×6, x y z, then rotation about X Y Z); GTSAM Pose3 tangent [ω; v] (gtsam/geometry/Pose3.cpp:100, right retraction T·Exp(ξ)); Sophus SE3 [υ; ω] (sophus/se3.hpp:840-850)GTSAM's order, rotation firstGTSAM's chart. Ours perturbs the rotation on the left, in the destination frame (VGC's RigidExpManifold6), the chart VisualGeometryToolkitEstimation reports in. The ROS layout is the same chart with the blocks swapped (ros_pose_covariance).
cameras/v6.0, COLMAPCOLMAP images.txt (doc/format.rst:143,153-163: world → camera, Hamilton w,x,y,z); pixel centres at 0.5 (doc/cameras.rst:52-57)Hamilton w,x,y,z; the COLMAP pixel frame as the canonical one—

Exporters

The exporters are write-only. The tests verify each by reading its text back (test/rig_schema_tests.jl).

  • save_tum_trajectory(path, tr; stamp = :time) writes TUM: metres, x,y,z,w, the pose of from in to, with a # comment header that evo skips.
  • save_euroc_trajectory(path, tr; stamp = :time) writes EuRoC ground truth: nanosecond timestamps, metres, w first. evo reads 8 or more columns.
  • save_bop_scene(to_frame, dir, tp, cameras; obj_ids, undistorted) writes scene_gt.json and scene_camera.json laid out as bop_toolkit's save_json writes them, with K moved into to_frame. It returns the obj_id and image maps.
  • save_trc(path, obs; data_rate, stamp) writes OpenSim TRC: tab-separated, in millimetres, with missing markers as empty fields.

There is no C3D writer. A maintained Python writer exists (ezc3d, on conda-forge), but a Python dependency for one binary format is out of proportion for this package, and TRC carries the same marker trajectories.