andrewadare / ecce

Extrinsic camera calibration example

ECCE: Extrinsic camera calibration example

Quickstart

To run this code, install the following dependencies:

• OpenCV
• GTSAM

Then follow a typical CMake workflow:

mkdir build && cd build
cmake ..
make

This builds any .cpp files in examples/ with static linkage to the library code in ecce/. They can then be run with no arguments. camera_to_vehicle_calibration saves several text files that can be used for analysis and visualization.

Plotting and graphics

Various python scripts are included in the python/ directory. Package dependencies can be installed (preferably into a dedicated environment) via pip install -r requirements.txt.

Note: file names written from camera_to_vehicle_calibration are hard-coded in scripts (TODO)

ecce's Issues

Add argparse to plotting scripts

The python scripts have hard-coded filenames.

Refactoring plan

If I were to refactor this project, I'd ditch the *Collection classes and instead define a CameraToVehicleGraph as the central data structure:

• Use int (or, I guess unsigned int) identifiers to key on objects instead of strings.
• Define a Pose3Node struct with id, pose, name.
• unordered_map<int, Pose3Node>: one for tags, one for cams. Key on node.id.
• Store intrinsics (or possibly a list of them) as members for camera construction.
• CameraToVehicleGraph would have getCameraNode(i), getTagNode(j)
• Define Side and TagZone enums.
• Define a CameraType enum for cameras, onboard vs external.
• Need a way to do something like getCameraId(CameraType, Side, index=-1).
  • Try lookup by defining a composite object as the key. This could be a struct or a tuple called CameraAttributes. Then do map<CameraAttributes, int> cameraLookup. Could also specialize a hash for CameraAttributes and use an unordered map. The latter would be a small internal optimization that could be implemented later without affecting the interfaces.
• Same story for getTagId(Side, TagZone).
• Define convenient interfaces to get camera/tag poses, names, etc. using the get id functions internally.
• pointMap would key on a pair<int, int>
• To iterate through graph:
  • Implement graph.edges(), which could be overloaded to query one side or all edges.
  • Make enums above iterable by defining a final NONE entry and checking for it.
• Don't use gtsam::Symbol class anywhere on the frontend. Just use the int keys.