dim_odom

Python bindings for dimSLAM: CuvslamOdometry wraps the cuVSLAM visual(-inertial) odometry front end, OdometryFusion the error-state Kalman filter that fuses odometry sources with an IMU. Everything runs in-process and on the caller's data stamps, so replaying a recording is deterministic.

import dim_odom

tracker = dim_odom.CuvslamOdometry(
    {"camera_mode": "stereo", "use_gpu": False},
    tf=lambda parent, child: ((0.0, 0.0, 0.0), (0.0, 0.0, 0.0, 1.0)),
)
tracker.handle_camera_info(dim_odom.CameraModel(...))
estimate = tracker.handle_image(dim_odom.ImageFrame(...))

fusion = dim_odom.OdometryFusion({"sources": {"odom->base_link": {}}})
if estimate is not None:
    fusion.handle_source(estimate)
fused = fusion.maybe_publish()

Transforms cross the boundary as ((x, y, z), (qx, qy, qz, qw)); the tf callable answers rigid-mount lookups (parent, child) with such a tuple or None.

The wheels bundle libcuvslam (open source) for each platform. Configs are dicts mirroring the Rust CuvslamOdometryConfig and OdometryFusionConfig; absent keys take the defaults.