perception-tools

Camera calibration and perception tools for CU Robotics' hive repository. Most tools must be run inside the calibration-tools docker service. Use a combination of the compose YAMLs when starting up the service:

• compose.x11.yaml: Accommodate X11 displays (for imshow and other GUIs). Open a PR if you use wayland.
• compose.nvidia.yaml: Mount an NVIDIA GPU for hardware acceleration. NVIDIA Container Toolkit required.

multi-camera calibration

The collection process for calibration images is documented in hive under the calibrate Cargo feature. Calibration batches should be placed under the hive-calib volume. Hive's pull_batch tool does this automatically.

$ python3 -m calibration_tools.calibrate -h

usage: calibrate [-h] -t TARGET [-b BATCH] -f FOCAL_ESTIMATE [-d]

process hive calibration batch

options:
  -h, --help            show this help message and exit
  -t, --target TARGET   target name
  -b, --batch BATCH     calibration batch name
  -f, --focal-estimate FOCAL_ESTIMATE
                        focal length estimate
  -d, --debug           enable debug features

TARGET must name a YAML under hive-calib that completely specifies the geometry of your calibration target. Generally we should opt to use strong_checkerboards. FOCAL_ESTIMATE specifies a preliminary estimate for the focal length in pixels and is required by the optimizer. You should use an existing calibration to seed this value.

All mrcal command-line tools are available within the container, including uncertainty and cross-validation utilities.

synthetic dataset creation

You can generate training or evaluation images of armor plates with deterministic poses using blender (installed under /opt/blender/blender). The synthetic dataset generation tools use mrcal camera models directly and should provide representative images of your camera rig. Blender uses a pinhole camera so renders are first remapped using your desired projection model. Generally a larger resolution than that of your physical camera is required to avoid excessive upsampling during this conversion, though you can modify the render resolution depending on performance needs.

$ ./scripts/synthesize_dataset --config <path-to-config>

All manifests for synthetic datasets must comply with the specifications in src/synthetic/common.py. An example manifest is given under configs/synthetic/default.json.